<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[The Tapeout Log]]></title><description><![CDATA[See the Unseen]]></description><link>https://thetapeoutlog.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Sat, 21 Nov 2026 18:05:50 GMT</lastBuildDate><atom:link href="https://thetapeoutlog.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[The Bridge, Not the Interposer: Microsoft Picks Intel's EMIB Packaging — This Week in Semiconductors]]></title><description><![CDATA[Microsoft is the first to announce Intel's EMIB packaging for its own cloud chips — plus a $10B fab breaks ground. What it means for PD engineers.]]></description><link>https://thetapeoutlog.hashnode.dev/the-bridge-not-the-interposer-microsoft-picks-intel-s-emib-packaging-this-week-in-semiconductors</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/the-bridge-not-the-interposer-microsoft-picks-intel-s-emib-packaging-this-week-in-semiconductors</guid><category><![CDATA[semiconductor]]></category><category><![CDATA[physical design]]></category><category><![CDATA[Advanced Packaging]]></category><category><![CDATA[EMIB]]></category><category><![CDATA[semiconductor news​]]></category><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Fri, 20 Nov 2026 03:58:39 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/f11bd827-4962-4377-8faf-86bbb48149fb.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><img src="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/f11bd827-4962-4377-8faf-86bbb48149fb.webp" alt="" /></p>
<p>This week the industry's story moved off the die and into the package. On November 18, Microsoft became the first company to publicly commit to using Intel Foundry's EMIB advanced packaging for the chips it designs for its own cloud, under a long-term agreement that also locks in multi-year supply of Intel's server processors. Five days earlier, GlobalFoundries broke ground on a new 300 mm fab backed by a government-funded investment initiative worth up to $21 billion — in which the funder provides the capital and the foundry provides the technology and runs the plant. And at its own annual conference, Microsoft unveiled its next AI accelerator and shipped its next in-house server CPU.</p>
<h2>What Microsoft and Intel actually announced</h2>
<p>Microsoft and Intel announced an expansion of their long-running collaboration under a long-term agreement covering select areas of the Azure cloud infrastructure. The headline is the packaging: Microsoft is the first company to announce use of Intel Foundry's EMIB (Embedded Multi-die Interconnect Bridge) advanced packaging technology for Azure first-party silicon — the chips Microsoft designs for its own cloud.</p>
<p>The agreement also confirms multi-year supply of Intel Xeon server processors to Azure, including the next generation of server CPUs, codenamed Diamond Rapids, plus collaboration on Compute Express Link (CXL) memory, Intel QuickAssist acceleration, and confidential computing. The announcement also quotes vendor test figures — up to 2.3x greater performance and memory flexibility in Azure M-series v4 virtual machines using CXL memory, and up to 2.5x faster SQL backups through the first QuickAssist-enabled virtual machines. Those are the vendors' own results, with a "results may vary" caveat, describing existing virtual machines — not the new packaging agreement itself.</p>
<h2>Why EMIB matters to a physical design engineer</h2>
<p>EMIB connects multiple dies side by side inside a single package using small silicon bridges embedded in the package's organic substrate, rather than a full silicon interposer underneath the dies. The bridge carries the high-density die-to-die links; the rest of the package stays on a conventional substrate.</p>
<p>That distinction is the whole point. A full interposer is a large piece of silicon that every die sits on, and a defect anywhere in it can scrap the assembly. Intel says EMIB turns more than 90% of a packaging wafer into usable product, against roughly 60% for interposer-based designs. The claim is Intel's, but the mechanism is straightforward: you only place small bridges where dies actually need to talk to each other, instead of paying for silicon under the entire package.</p>
<p>For anyone who does floorplanning and signoff, this is the direction the work is moving: the hardest interconnect problem is increasingly between dies — bump pitch, bridge routing, die-to-die timing, known-good-die testing, thermal coupling — not inside one die. The die is still where timing closes, but the package is where the system is won or lost.</p>
<h2>A first announcement is not a wafer commitment</h2>
<p>Keep the deal in proportion. Neither company disclosed the contract value, the volumes, the names of the Azure chips that will use EMIB, or when EMIB-packaged parts will appear in the fleet. The announcement concerns advanced packaging and processor supply, which is different from a commitment to manufacture a leading-edge chip on Intel's wafers. Microsoft's newest in-house chips are currently made by another foundry, and nothing in this release changes that. What Microsoft has bought is packaging capacity and Xeon supply — real, but not a leading-edge wafer order.</p>
<p>Intel's chief executive also said this week that Intel has signed its first external customer for its 18A manufacturing process — but did not name the client, so the size and timing of any production commitment are unknown. Until a customer name, a product, or a volume appears, EMIB is the verifiable part of the week.</p>
<h2>The other headline: shovels in the ground for a $10 billion fab</h2>
<p>On November 13, GlobalFoundries broke ground on Fab 8.2, a new 300 mm fab and the first facility under an investment initiative worth up to $21 billion. The first fab carries an initial investment of about $10 billion, supported by $570 million in public incentives, and is expected to come online in 2029. The new building will hold about 386,000 square feet of cleanroom, and the project is expected to create more than 1,100 manufacturing jobs.</p>
<p>The structure is the interesting part. Under the initiative, the funding partner provides the investment while GlobalFoundries contributes its technology and manufacturing expertise and operates the fab. A second fab of similar scale follows under the same initiative, and the capacity targets automotive, industrial, and connected-infrastructure chips.</p>
<p>A groundbreaking is not a wafer. Three years separate the ceremony from first production, and the economics of the fab will be decided by the utilisation and pricing of that capacity in 2029, not by the size of the announcement. But the model itself is new: a foundry operating a fab it did not pay to build, for a funder that does not make chips. If it works, expect it to be copied.</p>
<h2>Microsoft's own silicon, in the same week</h2>
<p>The EMIB announcement landed a day after Microsoft's annual Ignite conference, where the company announced Maia 300, its next-generation AI accelerator — which it says will deliver 40% more tokens per watt compared with leading GPU-based systems, the company's own claim, with no availability date given — and general availability of its next-generation Cobalt 200 virtual machines, built on its in-house Arm-based server CPU. Microsoft is designing more of its own silicon, and it is now also choosing who packages it. The chip, the package, and the server are becoming one co-design problem.</p>
<h2>What to watch next</h2>
<p>Three things will show which way this week goes. First, whether Microsoft names the Azure chip that will use EMIB — Cobalt, Maia, or both — and when it ships. Second, whether the unnamed 18A customer is revealed, and whether that commitment extends to wafers rather than packaging. And third, whether the GlobalFoundries fab stays on its 2029 schedule, and whether its operating model — funder pays, foundry runs — becomes a template for the next round of fab building.</p>
<p>For now, the practical read is simple. The most consequential interconnect in a modern system may no longer be a metal layer on a die, but a small silicon bridge buried in a package substrate. This week, the largest cloud buyer in the business agreed.</p>
<h2>Sources</h2>
<ul>
<li>Intel Newsroom, "Microsoft and Intel deepen collaboration to deliver differentiated Azure infrastructure," November 18, 2026 — <a href="https://www.intel.com/content/www/us/en/newsroom/news/corporate/microsoft-azure-infrastructure-collaboration.html">https://www.intel.com/content/www/us/en/newsroom/news/corporate/microsoft-azure-infrastructure-collaboration.html</a></li>
<li>The Inference, "Microsoft picks Intel Foundry's EMIB packaging for its own Azure AI chips," November 19, 2026 — <a href="https://theinference.org/article/microsoft-picks-intel-foundry-s-emib-packaging-for-its-own-azure-ai-chips">https://theinference.org/article/microsoft-picks-intel-foundry-s-emib-packaging-for-its-own-azure-ai-chips</a></li>
<li>XenoSpectrum, "Microsoft to Use Intel's EMIB Packaging for Custom Azure Chips in Multiyear Deal," November 18, 2026 — <a href="https://xenospectrum.com/en/azure-intel-emib-xeon-agreement/">https://xenospectrum.com/en/azure-intel-emib-xeon-agreement/</a></li>
<li>GlobalFoundries (company press release), "GlobalFoundries breaks ground on first new semiconductor fab," November 13, 2026 — <a href="https://gf.com/news-and-events/news/globalfoundries-breaks-ground-on-first-new-semiconductor-fab-under-landmark-us-japan-initiative/">https://gf.com/news-and-events/news/globalfoundries-breaks-ground-on-first-new-semiconductor-fab-under-landmark-us-japan-initiative/</a></li>
<li>Empire State Development, "Governor Hochul Celebrates the Groundbreaking of GlobalFoundries' New $10 Billion Semiconductor Manufacturing Facility," November 13, 2026 — <a href="https://esd.ny.gov/esd-media-center/press-releases/governor-hochul-announces-groundbreaking-globalfoundries-new-10-billion-semiconductor">https://esd.ny.gov/esd-media-center/press-releases/governor-hochul-announces-groundbreaking-globalfoundries-new-10-billion-semiconductor</a></li>
<li>Microsoft Blog, "Ignite 2026: Building the frontier ecosystem," November 17, 2026 — <a href="https://blogs.microsoft.com/blog/2026/11/17/ignite-2026-building-the-frontier-ecosystem/">https://blogs.microsoft.com/blog/2026/11/17/ignite-2026-building-the-frontier-ecosystem/</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[Clock Domain Crossing in Physical Design: What the Layout Has to Do After RTL Signs Off CDC]]></title><description><![CDATA[CDC in physical design: synchronizers, Gray-code FIFOs, STA exceptions, layout and scan — what PD must protect after RTL signs off CDC.]]></description><link>https://thetapeoutlog.hashnode.dev/clock-domain-crossing-in-physical-design-what-the-layout-has-to-do-after-rtl-signs-off-cdc</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/clock-domain-crossing-in-physical-design-what-the-layout-has-to-do-after-rtl-signs-off-cdc</guid><category><![CDATA[sta]]></category><category><![CDATA[physical design]]></category><category><![CDATA[cdc]]></category><category><![CDATA[#vlsi]]></category><category><![CDATA[Clock Domain Crossing]]></category><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Tue, 17 Nov 2026 04:49:44 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/c91ee7db-fe67-4ea2-8033-60465c42c078.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><img src="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/c91ee7db-fe67-4ea2-8033-60465c42c078.webp" alt="Clock Domain Crossing in Physical Design" /></p>
<p>Clock Domain Crossing in Physical Design: What the Layout Has to Do After RTL Signs Off CDC</p>
<p><em>Pradeep Kumar Adusumalli</em></p>
<p>A modern chip has many clocks, and every signal that moves between two of them crosses a boundary that static timing analysis cannot simply time its way across. The RTL team can sign off clock domain crossing with a clean CDC report, and the design can still fail in silicon — not because the RTL was wrong, but because physical implementation quietly broke the assumptions the RTL relied on. A synchronizer that is correct on paper stops being correct when synthesis separates its two flops, when placement puts them far apart, or when a timing exception hides the path from every tool that could have caught the damage. This post is about that second half of CDC.</p>
<h2>1. CDC in one picture</h2>
<p>Inside one clock domain, every flip-flop samples its input on the same clock edge, and static timing analysis guarantees the input is stable around that edge. Between two unrelated clocks there is no such guarantee. A signal from domain A can change at any moment relative to a clock edge in domain B, including right inside the setup-and-hold window of the receiving flip-flop. When it does, the receiving flop can go metastable: its output hovers between 0 and 1 for an unpredictable time before settling to one value or the other.</p>
<p>Metastability cannot be eliminated. It can only be given time to resolve. That is the whole idea behind the two-flop synchronizer, and it is also why CDC is a physical design problem and not only an RTL one: the time a metastable flop gets to resolve is a physical quantity, set by the clock period minus the delay of the wire and logic between the two flops. Anything that eats into that time — a long route, a buffer inserted by the tool, a slow cell chosen by optimization — eats directly into the safety of the crossing.</p>
<p>There is a second, separate problem that the synchronizer does not solve at all. Metastability is about one bit being uncertain. Coherence is about several bits meaning something together. A synchronizer protects one bit against one flop's uncertainty; it does not coordinate one bit with another. When two bits are supposed to mean something together, synchronizing each one correctly and separately can still produce a value the source never sent. That distinction — metastability versus coherence — decides which crossing structure to use, and it is the mistake that survives most often into silicon, because each individual crossing looks correct.</p>
<p>CDC bugs pass simulation, because ordinary RTL simulation does not model metastability, and they appear only now and then on silicon. That is why teams check CDC structurally, with tools that read the design and verify every crossing, rather than relying on simulation coverage.</p>
<h2>2. Why RTL-level CDC signoff is not enough</h2>
<p>RTL CDC signoff answers a structural question: does every crossing have a recognized synchronizer, handshake, or FIFO on it? Tools such as SpyGlass CDC, Questa CDC and JasperGold CDC perform that analysis on the elaborated netlist, with no stimulus, finding every net whose driver and receiver sit in different asynchronous clock domains and checking the structure on the path. That is powerful, and it is necessary. It is not sufficient, because the RTL signoff is done on a netlist that physical implementation is about to change.</p>
<p>Synthesis can retime, duplicate, or merge the flops of a synchronizer unless it is told not to. Placement can put the two flops far apart, so the wire between them consumes the resolution time the second flop exists to provide. Clock tree synthesis gives the two flops real, different clock arrival times, and scan insertion replaces flops and adds test logic around them. Any one of these steps can turn a synchronizer that passed CDC lint at RTL into a structure that no longer matches what the lint tool checked. The physical design job is to preserve the crossing structure through implementation, and to constrain it so the tools help rather than hurt. The RTL decides the structure. PD decides whether the structure survives.</p>
<h2>3. The two-flop synchronizer, and what PD must protect</h2>
<p>The standard fix for a single, slow-changing bit — a status level, an enable, a mode bit — is two flip-flops in the destination clock domain, back to back, with no logic between them. The first flop may go metastable, but it has a whole destination-clock cycle to settle before the second flop samples it.</p>
<p><strong>No logic between the two flops.</strong> Any combinational logic between them consumes resolution time, and a synthesis tool that is not told otherwise will happily place some there. The input must also come directly from a flip-flop in the source domain, not from combinational logic: logic in the source domain can glitch as its inputs change, and the destination clock is unrelated and can sample a glitch as a real pulse. A runt pulse on the wire can likewise drive the first flop into a deeper metastable state than the MTBF calculation assumes, because that calculation assumes a sharp, clean edge.</p>
<p><strong>Keep the two flops close together.</strong> The path between them must be short and lightly loaded. A long route adds wire delay and capacitance, and gives the implementation tool a reason to insert buffers on the path, each consuming resolution time. Mark the synchronizer flops so the tools preserve the structure: tool-specific attributes, a don't-touch on the pair, or a dedicated synchronizer cell from the library, some of which are characterized for faster metastability resolution. Constrain the path too, so the tool has an incentive to keep it short (section 7).</p>
<p><strong>Do not let the tools optimize, retime, or duplicate the first flop.</strong> The first flop must have no fan-out except to the second flop. If its output fans out to other logic, that logic sees a value that may still be metastable. Marking the synchronizer and checking the CDC report after synthesis — not only at RTL — is how this gets caught.</p>
<p><strong>Choose the number of stages by calculation, not by habit.</strong> MTBF grows exponentially with the resolution time the first flop is given, and a third flop adds a whole clock period of resolution time. At moderate clock rates two flops typically give an MTBF of many years or more; at very high clock rates, designs use three flops. The failure rate can be made extremely small and can never be made zero.</p>
<p><strong>A synchronizer transfers a level, not an event.</strong> A one-cycle pulse from a fast domain can fall entirely between two edges of a slow clock and be missed altogether. The fix is a toggle synchronizer: it flips a level in the source domain once per event, the level crosses through a two-flop synchronizer, and the destination detects the edge with an XOR of the last two synchronized samples. The third flop there is a delayed copy for edge detection, not a third synchronizer stage. Source events must be spaced at least two or three destination clock cycles apart, or two toggles can cancel and an event is lost silently. Faster, back-to-back events need a handshake or a FIFO.</p>
<h2>4. Multi-bit data: Gray code, handshakes, and asynchronous FIFOs</h2>
<p>If each bit of a bus is synchronized separately, each bit resolves independently. When the bus changes from 0111 to 1000, the receiver might see any mix of old and new bits for a cycle — a word the source never held. No amount of synchronization fixes that, because the problem is coherence, not metastability. There are three safe structures, and the choice between them is a choice about what the destination must learn.</p>
<p><strong>Gray-coded pointers, for counters that step by one.</strong> In Gray code, consecutive values differ in exactly one bit, so a pointer sampled while changing gives either the old value or the new value — both valid. The asynchronous FIFO is built on this: the data sits in a dual-port RAM written in one domain and read in the other, and only the read and write pointers cross between domains, as Gray code through two-flop synchronizers. The conversion order is the whole correctness of the pattern: increment as binary or Gray, cross as Gray, and convert back to binary only after the second synchronizer flop, in the destination domain. The synchronized pointer may lag by a cycle or two, so the FIFO sometimes reports full or empty slightly early — always the safe direction. The guarantee holds only for a value that steps by one, monotonically. A counter that can jump, be loaded, or be cleared to zero breaks it immediately. Gray code is the pointer solution, not a general multi-bit solution.</p>
<p><strong>The four-phase handshake, for an occasional multi-bit word.</strong> The source puts the data on a bus and holds it stable, then raises a request. The request is synchronized into the destination, which samples the now-stable data and raises an acknowledge. The acknowledge is synchronized back to the source, which only then changes the data. The four phases are req rises, ack rises, req falls, ack falls. Only the single-bit req and ack signals are synchronized; the data bus itself is never synchronized bit by bit. The cost is speed: a handshake takes several cycles per word — roughly four synchronization delays, typically around eight to ten cycles in total, and in one worked example across a 10 ns and a 14 ns clock, each word took 140 ns. For a configuration word written once at boot that is irrelevant; for a data stream it is ruinous.</p>
<p><strong>The asynchronous FIFO, for a stream.</strong> The FIFO decouples the two rates as well as the two clocks: the write side fills a small memory in one domain, the read side empties it in the other, and only the Gray-coded pointers cross. It sustains up to one word per cycle of the slower clock, at the cost of a memory and the Gray-pointer machinery. That is the structure for a continuous stream of data, and the reason bridges to peripheral buses usually use a handshake while clock converters on data paths use FIFOs. One more option is worth naming because it is the cheapest crossing of all: recompute the value in the destination domain instead of crossing it. The best crossing is no crossing.</p>
<h2>5. Reconvergence: the bug that survives correct synchronizers</h2>
<p>The most dangerous CDC failure in physical design is the one where every individual crossing is correct. Two related signals — an enable and a mode bit, say — each get their own proper two-flop synchronizer. Each synchronizer is individually correct. But metastability resolution is per-flop and independent, so on the cycle the signals change, one may resolve one destination cycle before the other. For that one cycle the destination sees a combination the source never sent, and logic that combines the two signals — logic that reconverges them — acts on a phantom value.</p>
<p>The fix is a rule: anything that must be interpreted together must cross together. Combine the signals into a single validation signal in the source domain and synchronize that one bit, or treat the pair as a multi-bit value and cross it with a handshake or a FIFO. CDC lint tools flag reconvergence as a violation, and it is worth reading the reconvergence section of the CDC report rather than only the unsynchronized-crossing list, because a clean unsynchronized-crossing count says nothing about this failure. The same hazard applies at the layout level: two synchronizers that are correct in RTL can be given very different routing delays by placement. Keeping related synchronizers in the same region, with matched routing, is part of the placement job.</p>
<h2>6. Resets are a crossing too</h2>
<p>Reset gets forgotten because it does not look like data. It is a signal that one domain drives and another domain's flops obey, which is the definition of a crossing. The standard answer is to assert asynchronously and deassert synchronously, in each clock domain separately: the raw reset asserts a chain of flops immediately, with no clock needed, and the release is clocked out of that chain by the local clock, so reset drops cleanly one or two edges after the source releases. Every clock domain needs its own reset synchronizer; sharing one across domains reintroduces exactly the crossing it was built to remove. If reset deasserts near a clock edge, some flops in a domain can leave reset one cycle before others, and a state machine can start in a state its encoding does not allow. Reset domain crossing is checked with dedicated RDC tools alongside CDC, and it deserves the same waiver discipline as any other crossing.</p>
<h2>7. STA: what to constrain, and what not to hide</h2>
<p>CDC paths must be excluded from normal setup and hold checks, because there is no meaningful timing relationship between two unrelated clocks — the arrival time of data relative to the capture clock is unbounded, and timing the path as if it were synchronous produces false violations. But "excluded from setup and hold" is not the same as "ignored", and the difference is where most CDC timing mistakes happen.</p>
<p>The cleaner way to declare the relationship is <code>set_clock_groups -asynchronous</code>, which states the intent — these clocks are unrelated — covers both directions in one command, and, unlike a bare <code>set_false_path</code>, also carries the signal-integrity meaning that an aggressor in an asynchronous domain can switch at any time relative to a victim. A <code>set_false_path</code> between the same clocks has the same effect on timing, but it says only that the path is false, which is not true: the path exists, and the command is only a mechanism for telling the tool not to time it. <code>set_clock_groups</code> is the preferred form for that reason.</p>
<p>The trap is that either form makes the path invisible to the tools. Once a path is a false path, nothing tells synthesis and place-and-route to keep it short, and the synchronizer flops can end up at opposite corners of the block with a long, buffered route between them — exactly the damage section 3 describes. The better practice, documented in the CDC timing-constraints literature, is to keep the CDC paths visible to the timing engine and constrain them with a <code>set_max_delay</code> that is just large enough for the path, with the value determined experimentally for the technology, the design, and sometimes the specific path. That gives the implementation tools an incentive to keep the synchronizer path short and lightly loaded. A <code>set_max_delay</code> relaxes only the setup and maximum-delay checks and leaves the hold and minimum-delay checks unaffected, so the hold side needs its own attention — a <code>set_min_delay</code> where the flow calls for it.</p>
<p>Two details are worth getting right. First, <code>set_clock_groups -asynchronous</code> has an <code>-allow_paths</code> option: with it, the tool does not apply false paths between the asynchronous domains, while still recording the asynchronous relationship — the form to use when the intent is to keep the CDC paths enabled and timed, rather than hidden. Second, the synchronizer's own internal path — from the first flop to the second — is a same-clock path in the destination domain and should be timed as one, with a full destination cycle available for resolution. The practical rule is simple to state and easy to violate: declare the clocks asynchronous, constrain the crossing paths so the tools keep them short, and never use an exception to hide a path nobody has reviewed. A timing exception hides a real path from STA; a CDC waiver hides a real crossing from CDC analysis.</p>
<h2>8. Layout: routing, shielding, and the multi-voltage crossing</h2>
<p>The synchronizer input is the most sensitive net in the block. It is the one net where a glitch is not a cosmetic problem but a functional one, because the destination clock can sample the glitch as a real pulse. Three layout habits protect it. Keep the route between the source flop and the first synchronizer flop short, and keep the route between the two synchronizer flops shorter still — the resolution budget is spent on wire before it is spent on anything else. Do not route the synchronizer input alongside a strong aggressor for a long parallel run; a crosstalk-induced glitch on that net is indistinguishable from a real transition to the receiving flop, and shielding or spacing is cheap insurance on a net that exists in small numbers. And do not let the synchronizer input take a long detour around a macro or through a congested channel that forces extra vias — a path that crosses a power-domain boundary without a level shifter where one is required is the same class of mistake.</p>
<p>When the two clock domains also sit in different voltage domains, CDC and power intent interact: the synchronizer flops belong in the destination domain, and the signal arriving at the first synchronizer flop must be a clean signal at the destination domain's voltage. Level shifters and isolation cells change the path delay and glitch behaviour, so CDC checks should be re-run after power-intent insertion, not only before it.</p>
<h2>9. Scan and DFT: test must not break the crossing</h2>
<p>Scan insertion is the step most likely to break a synchronizer silently. A scan chain that stitches flip-flops across asynchronous domains mixes clocks in a single chain, and the shift path between two domains is a CDC path that test mode exercises every cycle. The standard handling is to avoid stitching across asynchronous domains: use separate scan chains per domain, or a single test clock muxed into the chain, with lock-up latches at the domain boundary to hold the shift data steady across the clock skew between domains. CDC paths are tested separately, using functional test or dedicated logic, rather than through the scan chain.</p>
<p>The same care applies to at-speed capture: a capture clock faster than a domain's functional clock can create setup pressure on a comfortable path, and a CDC path that is a false path functionally must not become a timing path in test mode by accident. Synthesis and STA need SDC for both functional and test modes, and the design must converge on timing in both. Run the CDC and STA reports in test mode, not only in functional mode, and confirm that no synchronizer has been restructured by scan insertion.</p>
<h2>10. The checklist</h2>
<p>Before signoff, walk the design with this list. Every item is a check that RTL signoff alone does not perform.</p>
<ul>
<li><strong>Every crossing is classified by kind, not by width.</strong> Level, event, counter, arbitrary multi-bit value, stream, reset. A one-bit pulse is more dangerous than a thirty-two-bit stable bus.</li>
<li><strong>Every single-bit crossing has a two-flop synchronizer with no logic between the flops, no fan-out from the first flop, and the input driven directly from a source-domain flop.</strong> Three flops where the resolution calculation calls for it.</li>
<li><strong>Every multi-bit crossing uses Gray code, a handshake, or an asynchronous FIFO — never per-bit synchronizers.</strong> Gray code only for values that step by one, monotonically, with the decode after the second flop.</li>
<li><strong>Related signals cross together.</strong> No reconvergence: anything interpreted together in the destination crosses through one structure.</li>
<li><strong>Every reset crossing has its own reset synchronizer per domain.</strong> Assert asynchronously, release synchronously.</li>
<li><strong>The synchronizer flops are marked, kept close together, and reviewed after synthesis and after placement.</strong> Check the CDC report again after each step, not only at RTL.</li>
<li><strong>The clocks are declared asynchronous with <code>set_clock_groups -asynchronous</code>, and the crossing paths are constrained with a <code>set_max_delay</code> that keeps them short — not hidden with a bare false path.</strong></li>
<li><strong>The synchronizer input routing is short, clean, and reviewed.</strong> No long parallel run beside an aggressor, no unnecessary detour, no missing level shifter where a voltage boundary is crossed.</li>
<li><strong>Scan chains do not cross asynchronous domains.</strong> Separate chains or a muxed test clock, lock-up latches at the boundary, and CDC and STA re-checked in test mode.</li>
<li><strong>Every waiver carries an argument.</strong> Why the crossing is safe, what property makes it safe, and what change would break it. A waiver that says only "known crossing" is a hole in the signoff, not a record of one.</li>
</ul>
<p>CDC is not a block that RTL finishes and PD inherits. It is a contract between the two: RTL chooses a structure whose correctness depends on physical properties — resolution time, wire delay, placement, routing — and PD is the part of the flow that decides whether those properties hold. When the contract is kept, a crossing is safe in a way that survives synthesis, placement, clock-tree synthesis, scan insertion, and power-intent insertion. When it is not kept, the failure shows up exactly where it is hardest to find: once in millions of crossings, on some boards, at some temperatures, in a chip that passed every simulation.</p>
<h2>Sources</h2>
<ul>
<li>ChipVerify — Synchronizers &amp; Clock Domain Crossing — <a href="https://chipverify.com/digital-fundamentals/synchronizers-clock-domain-crossing">https://chipverify.com/digital-fundamentals/synchronizers-clock-domain-crossing</a></li>
<li>VLSI Web — Clocking and Synchronisation in VLSI: CDC Made Simple — <a href="https://vlsiweb.com/clocking-and-synchronization-in-vlsi/">https://vlsiweb.com/clocking-and-synchronization-in-vlsi/</a></li>
<li>ChipVerify — Bridges and Clock Domain Crossing — <a href="https://chipverify.com/bus-protocols/bridges-and-clock-domain-crossing">https://chipverify.com/bus-protocols/bridges-and-clock-domain-crossing</a></li>
<li>VLSI Web — Handshake Synchronisers for CDC: Four-Phase and Two-Phase Explained — <a href="https://vlsiweb.com/handshake-protocols/">https://vlsiweb.com/handshake-protocols/</a></li>
<li>VLSI Mentor — Gray-Code Pointer Synchronization — <a href="https://www.vlsimentor.com/rtl-design-patterns/gray-pointer-sync">https://www.vlsimentor.com/rtl-design-patterns/gray-pointer-sync</a></li>
<li>VLSI Mentor — CDC Design Rules &amp; Lint — <a href="https://www.vlsimentor.com/rtl-design-patterns/cdc-design-rules">https://www.vlsimentor.com/rtl-design-patterns/cdc-design-rules</a></li>
<li>Lizard-Spock — SoC Intermediate 05: Clock Domain Crossing Techniques — <a href="https://lizard-spock.co.uk/soc-intermediate-05-clock-domain-crossing-techniques.html">https://lizard-spock.co.uk/soc-intermediate-05-clock-domain-crossing-techniques.html</a></li>
<li>Brabec — Timing Constraints for Clock-Domain Crossings (GitHub gist) — <a href="https://gist.github.com/brabect1/7695ead3d79be47576890bbcd61fe426">https://gist.github.com/brabect1/7695ead3d79be47576890bbcd61fe426</a></li>
<li>AMD Support — Constraining Asynchronous Clocks — <a href="https://adaptivesupport.amd.com/s/question/0D52E00007FSg1JSAT/constraining-asynchronous-clocks?language=zh_CN">https://adaptivesupport.amd.com/s/question/0D52E00007FSg1JSAT/constraining-asynchronous-clocks?language=zh_CN</a></li>
<li>Cummings — Simulation and Synthesis Techniques for Asynchronous FIFO Design, SNUG San Jose 2002 — <a href="http://www.gstitt.ece.ufl.edu/courses/fall19/eel4720_5721/reading/CummingsSNUG2002SJ_FIFO2.pdf">http://www.gstitt.ece.ufl.edu/courses/fall19/eel4720_5721/reading/CummingsSNUG2002SJ_FIFO2.pdf</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[The PD Engineer's Reading List: Papers and Books That Actually Teach the Craft]]></title><description><![CDATA[Pradeep Kumar Adusumalli Most reading lists for physical design are link dumps: fifty titles, no opinions, no way to tell which are worth your evening…]]></description><link>https://thetapeoutlog.hashnode.dev/the-pd-engineer-s-reading-list-papers-and-books-that-actually-teach-the-craft</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/the-pd-engineer-s-reading-list-papers-and-books-that-actually-teach-the-craft</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Sun, 15 Nov 2026 04:51:40 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/3910cbbb-60b7-40c6-862b-a2cbb24afdf1.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><img src="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/3910cbbb-60b7-40c6-862b-a2cbb24afdf1.webp" alt="" /></p>
<p><em>Pradeep Kumar Adusumalli</em></p>
<p>Most reading lists for physical design are link dumps: fifty titles, no opinions, no way to tell which are worth your evening. This one is twelve entries long, every one verified to exist with real authors and a real link, and every one chosen because it teaches something you will actually use on a real design. No tool manuals — those change every release, and a manual teaches you buttons, not judgment.</p>
<p>Two ways to use it. New to PD? Read it in order — the first book gives you the whole flow, and the rest fill in one corner each. Already working? Pick by problem: timing trouble, start with entry 2; a router that will not converge, entry 10; a clock tree that will not close, entry 12. One warning up front: reading alone does not make a PD engineer. Each of these earns its place only when you go back to your design afterwards and see the idea working there.</p>
<h2>The books</h2>
<h3>1. VLSI Physical Design: From Graph Partitioning to Timing Closure — Andrew Kahng, Jens Lienig, Igor Markov, Jin Hu (2nd edition, 2022)</h3>
<p>If you read one book on this list, read this one. It walks the whole flow in a single volume: partitioning, chip planning, placement, global and detailed routing, and timing closure — and it explains the algorithms behind each step instead of just describing the step. That is the point. When you understand why the placer minimizes wirelength the way it does, the tool's behaviour stops looking random. For a fresher it is the fastest map of the field; for a working engineer it explains the why behind the flow you run every day. It is algorithms-first, so do not expect it to teach you any specific tool's commands — that is what the tool documentation is for.</p>
<p><a href="https://link.springer.com/book/10.1007/978-3-030-96415-3">VLSI Physical Design: From Graph Partitioning to Timing Closure</a></p>
<h3>2. Static Timing Analysis for Nanometer Designs: A Practical Approach — J. Bhasker and Rakesh Chadha (2009)</h3>
<p>This is the STA book. It starts where most timing introductions stop: cell timing and power modelling, interconnect delay, crosstalk and noise, and then the full verification of a chip with static timing analysis — on-chip variation, clock gating, half-cycle paths, and the timing of source-synchronous interfaces. It is written for working engineers, and it reads that way: practical, dense, and organised around the problems you actually hit in signoff. One honest note: it is from 2009, so some of the node-specific numbers are dated. The method is not. If you sign off timing, this belongs within reach of your desk.</p>
<p><a href="https://doi.org/10.1007/978-0-387-93820-2">Static Timing Analysis for Nanometer Designs</a></p>
<h3>3. CMOS VLSI Design: A Circuits and Systems Perspective — Neil Weste and David Harris (4th edition, 2010)</h3>
<p>The broad foundation. It starts at transistor operation and CMOS gates, works through layout, timing and low power, and ends at full systems-on-chip. Beginners get a book that stays readable while going deep; experienced engineers get a reference for the "how does this actually work" question. The fourth edition is from 2010, so treat its newest process details as history; the circuits, the layout thinking and the design judgment have not aged.</p>
<p><a href="https://openlibrary.org/books/OL17983609M/CMOS_VLSI_design">CMOS VLSI Design: A Circuits and Systems Perspective</a></p>
<h3>4. Digital Integrated Circuits: A Design Perspective — Jan Rabaey, Anantha Chandrakasan, Borivoje Nikolic (2nd edition, 2002)</h3>
<p>Rabaey's book bridges the two views every PD engineer needs: the circuit view (what the inverter and the flop actually do) and the system view (what timing, clocking, power and interconnect do to a design). Its chapters on the CMOS inverter, sequential logic, timing and clocking, and design for low power are the reason it is still recommended two decades after publication. Read it to understand what the standard cells in your netlist are really doing — that understanding separates an engineer who runs the flow from one who can debug it. Some of the process data is old; the reasoning is not.</p>
<p><a href="https://www.booksamillion.com/p/Digital-Integrated-Circuits/Jan-Rabaey/9780130909961">Digital Integrated Circuits: A Design Perspective</a></p>
<h3>5. Handbook of Algorithms for Physical Design Automation — edited by Charles Alpert, Dinesh Mehta and Sachin Sapatnekar (2008)</h3>
<p>This is not a book to read cover to cover. It is a reference of more than a thousand pages, written by specialists, covering the algorithms of the whole physical design flow in depth. You open it when you need to understand something properly: how a particular placement or routing algorithm works, what its limits are, and what the alternatives are. It is the most demanding book on this list, and the most useful once you are past the beginner stage. Keep it as the second opinion you consult when the tool does something you cannot explain.</p>
<p><a href="https://www.google.com.br/books/edition/Handbook_of_Algorithms_for_Physical_Desi/w1Qy4fbtCkMC">Handbook of Algorithms for Physical Design Automation</a></p>
<h3>6. Low-Power CMOS Design — edited by Anantha Chandrakasan and Robert Brodersen (1998)</h3>
<p>A collection of landmark papers on low power, edited into one volume: component technologies and circuits, architecture, system design, and CAD techniques. It is the right book when your problem is power — leakage, dynamic power, power gating, or the power implications of a physical design choice — because it shows how the field's best ideas fit together instead of presenting one technique in isolation. Read it as a collection, not a textbook: pick the chapters that match the problem in front of you. It is from 1998, so it predates the multi-voltage, power-gating-heavy flows of today, but the fundamentals of where power goes and how to attack it are all here.</p>
<p><a href="https://doi.org/10.1109/9780470545058">Low-Power CMOS Design</a></p>
<h3>7. Physical Design Automation of VLSI Systems — edited by Bryan Preas and Michael Lorenzetti (1988)</h3>
<p>The classic. Published in 1988, it covers partitioning, placement, floorplanning, routing, layout verification and the complexity of layout problems — the same problems, in the same order, that a modern flow still solves. The machines changed beyond recognition; the problems did not. Read it for the thinking: how the pioneers framed placement as an optimization problem, why routing is hard, where the heuristics came from. The numbers and the tool descriptions are history. The framing is still the clearest version of it anywhere, and it is the book that makes the newer literature easier to read.</p>
<p><a href="https://books.google.tn/books?id=GyZTAAAAMAAJ">Physical Design Automation of VLSI Systems</a></p>
<h2>The papers</h2>
<h3>8. W. C. Elmore, "The Transient Response of Damped Linear Networks with Particular Regard to Wideband Amplifiers" (Journal of Applied Physics, 1948)</h3>
<p>Nine pages, and almost eighty years later every delay model in your flow still descends from it. Elmore showed how to define delay time and rise time from the Laplace system function of a network — the idea that became Elmore delay, the first-order RC delay estimate used in timing analysis, placement and routing to this day. Read the original, not a summary: it is short and readable, and it shows you exactly what the model assumes — which is also what it gets wrong. Anyone who works with interconnect delay should know where the number in the timing report comes from.</p>
<p><a href="https://pubs.aip.org/aip/jap/article-abstract/19/1/55/159107/The-Transient-Response-of-Damped-Linear-Networks">Journal of Applied Physics</a></p>
<h3>9. C. M. Fiduccia and R. M. Mattheyses, "A Linear-Time Heuristic for Improving Network Partitions" (Design Automation Conference, 1982)</h3>
<p>Every placer starts by partitioning, and this is the paper that made partitioning fast. Fiduccia and Mattheyses took the Kernighan–Lin heuristic and made each pass run in time linear in the size of the netlist: move one vertex at a time instead of swapping pairs, work directly on hypergraphs instead of plain graphs, and keep vertices in gain buckets so the best move is always at hand. The result is the FM algorithm, still the backbone of partitioning in modern tools. Read it to understand what your placer is doing in its first minutes — and why a bad partition early on is so expensive to fix later.</p>
<p><a href="https://doi.org/10.1109/DAC.1982.1585498">Design Automation Conference</a></p>
<h3>10. C. Y. Lee, "An Algorithm for Path Connections and Its Applications" (IRE Transactions on Electronic Computers, 1961)</h3>
<p>The maze router. Lee's algorithm floods a grid with a wave from the source and traces the shortest path back — breadth-first search, in two phases called wave propagation and retrace. It guarantees the shortest path if one exists, and it is the ancestor of every detailed router you have ever run. It is slow and memory-hungry on a large grid, which is why its improvements — better labelling, guided search, rip-up and reroute — fill the routing literature. Read it when the router is taking forever and you want to know what it is actually doing, and why there is no free lunch in routing.</p>
<p><a href="https://doi.org/10.1109/TEC.1961.5219222">IRE Transactions on Electronic Computers</a></p>
<h3>11. S. Kirkpatrick, C. D. Gelatt and M. P. Vecchi, "Optimization by Simulated Annealing" (Science, 1983)</h3>
<p>The most readable paper on this list, and the one with the widest reach. Kirkpatrick and his co-authors drew the analogy between the slow cooling of a metal and the search for a good solution to a combinatorial problem: accept a worse move sometimes, with a probability that falls as the "temperature" drops, and you can escape the local minima that trap greedy methods. Simulated annealing went on to power placement, routing and optimization across the whole flow — and it explains something every PD engineer has seen: why the tool's result changes from run to run, and why a longer run usually finds a better answer. Read it for the idea, not the maths; the idea is the whole paper.</p>
<p><a href="https://doi.org/10.1126/science.220.4598.671">Science</a></p>
<h3>12. Phillip Restle and colleagues, "A Clock Distribution Network for Microprocessors" (IEEE Journal of Solid-State Circuits, 2001)</h3>
<p>The clocking paper. Restle describes the global clock strategy used on real microprocessor chips: buffered, tunable trees feeding a single common grid across most of the die, with measured global skew as low as 22 picoseconds on large chips. It is the clearest real-silicon account of the tree-versus-grid trade-off: trees are cheap and controllable, grids are low-skew and power-hungry, and the winning answer is a hybrid. Read it before your next clock tree synthesis run: it will change how you think about skew budgets, and it is the paper to cite when someone asks why the clock network costs so much power.</p>
<p><a href="https://doi.org/10.1109/4.918917">IEEE Journal of Solid-State Circuits</a></p>
<h2>How to actually use this list</h2>
<p>Do not collect these; read them. One paper understood beats ten bookmarked. Pick the entry that matches this week's problem, read it, then go back to your design and find the idea in it — recompute an Elmore delay by hand for one of your own nets, or code up the FM gain-bucket idea on a small netlist over a weekend. That is the step that turns reading into craft.</p>
<p>One honest gap in the list: there is no single great book on design for manufacturability here. DFM lives in foundry rule decks, process documentation and the layout rules you check every day — no textbook replaces reading those. And if you have a book or paper that taught you more than anything on this list, that is the real test of a reading list: it should start arguments, not end them.</p>
<h2>Sources</h2>
<ul>
<li>VLSI Physical Design: From Graph Partitioning to Timing Closure — <a href="https://link.springer.com/book/10.1007/978-3-030-96415-3">https://link.springer.com/book/10.1007/978-3-030-96415-3</a></li>
<li>Static Timing Analysis for Nanometer Designs: A Practical Approach — <a href="https://doi.org/10.1007/978-0-387-93820-2">https://doi.org/10.1007/978-0-387-93820-2</a></li>
<li>CMOS VLSI Design: A Circuits and Systems Perspective — <a href="https://openlibrary.org/books/OL17983609M/CMOS_VLSI_design">https://openlibrary.org/books/OL17983609M/CMOS_VLSI_design</a></li>
<li>Digital Integrated Circuits: A Design Perspective — <a href="https://www.booksamillion.com/p/Digital-Integrated-Circuits/Jan-Rabaey/9780130909961">https://www.booksamillion.com/p/Digital-Integrated-Circuits/Jan-Rabaey/9780130909961</a></li>
<li>Handbook of Algorithms for Physical Design Automation — <a href="https://www.google.com.br/books/edition/Handbook_of_Algorithms_for_Physical_Desi/w1Qy4fbtCkMC">https://www.google.com.br/books/edition/Handbook_of_Algorithms_for_Physical_Desi/w1Qy4fbtCkMC</a></li>
<li>Low-Power CMOS Design — <a href="https://doi.org/10.1109/9780470545058">https://doi.org/10.1109/9780470545058</a></li>
<li>Physical Design Automation of VLSI Systems — <a href="https://books.google.tn/books?id=GyZTAAAAMAAJ">https://books.google.tn/books?id=GyZTAAAAMAAJ</a></li>
<li>Journal of Applied Physics — Elmore, "The Transient Response of Damped Linear Networks" — <a href="https://pubs.aip.org/aip/jap/article-abstract/19/1/55/159107/The-Transient-Response-of-Damped-Linear-Networks">https://pubs.aip.org/aip/jap/article-abstract/19/1/55/159107/The-Transient-Response-of-Damped-Linear-Networks</a></li>
<li>Design Automation Conference — Fiduccia and Mattheyses, "A Linear-Time Heuristic for Improving Network Partitions" — <a href="https://doi.org/10.1109/DAC.1982.1585498">https://doi.org/10.1109/DAC.1982.1585498</a></li>
<li>IRE Transactions on Electronic Computers — Lee, "An Algorithm for Path Connections and Its Applications" — <a href="https://doi.org/10.1109/TEC.1961.5219222">https://doi.org/10.1109/TEC.1961.5219222</a></li>
<li>Science — Kirkpatrick, Gelatt and Vecchi, "Optimization by Simulated Annealing" — <a href="https://doi.org/10.1126/science.220.4598.671">https://doi.org/10.1126/science.220.4598.671</a></li>
<li>IEEE Journal of Solid-State Circuits — Restle and colleagues, "A Clock Distribution Network for Microprocessors" — <a href="https://doi.org/10.1109/4.918917">https://doi.org/10.1109/4.918917</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[Hold Time: The Violation That Fails at Every Corner]]></title><description><![CDATA[Pradeep Kumar Adusumalli Setup violations get all the attention in timing closure. They carry the big negative numbers, they decide your maximum clock…]]></description><link>https://thetapeoutlog.hashnode.dev/hold-time-the-violation-that-fails-at-every-corner</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/hold-time-the-violation-that-fails-at-every-corner</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Sat, 14 Nov 2026 05:57:00 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/73868764-cb9d-4994-ac92-8aa947a4b82a.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><img src="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/73868764-cb9d-4994-ac92-8aa947a4b82a.webp" alt="Hold time" /></p>
<p><em>Pradeep Kumar Adusumalli</em></p>
<p>Setup violations get all the attention in timing closure. They carry the big negative numbers, they decide your maximum clock frequency, and the whole team watches them. Hold violations are quieter, smaller, and far more dangerous. A setup violation means your chip runs slower than planned. A hold violation means your chip does not work at all — at any frequency, in any mode. And unlike setup, hold does not stay in one corner. It shows up wherever the fast side of your design is fastest, and at modern nodes that can be almost anywhere.</p>
<p>This post is about what hold time actually is, why it fails at every corner, why it is more dangerous than setup once silicon exists, and how it gets fixed in practice.</p>
<h2>1. Hold time, in one picture</h2>
<p>A flip-flop has two timing windows around its clock edge. Setup time is the window <em>before</em> the edge: the data must arrive and settle before the edge, or the flop samples something that has not finished changing. Hold time is the window <em>after</em> the edge: the data must stay put for a short time after the edge, or the new value races in and overwrites the value the flop is in the middle of capturing.</p>
<p>In a flop-to-flop path, that gives the hold check:</p>
<p>Hold slack = (clock-to-Q delay + logic delay) − (hold time + clock skew + uncertainty)</p>
<p>Every term in that equation is a <em>minimum</em> delay — the fastest the data can possibly arrive. Setup is the opposite check: it uses maximum delays and asks whether data arrives before the <em>next</em> edge. Hold uses minimum delays and asks whether data stays away until after the <em>same</em> edge.</p>
<p>The most important thing about the hold equation is what is missing from it: the clock period. It never appears. Hold is a same-edge race, so the time between edges is irrelevant. A path that fails hold fails at 1 GHz and at 10 MHz and at 1 Hz. Slowing the clock — the first thing anyone reaches for when timing is bad — does nothing at all.</p>
<h2>2. The fast corner is where hold lives — and it is not the only place</h2>
<p>In classic CMOS, transistors are fastest when the process is fast, the voltage is high, and the temperature is low. So the textbook rule is simple: close setup at the slow corner (slow process, low voltage, high temperature) and close hold at the fast corner (fast process, high voltage, low temperature). That rule is still the starting point, but it is no longer the whole truth, for three reasons.</p>
<p><strong>First, the fast corner is not one corner.</strong> A real signoff runs multi-mode multi-corner analysis — functional mode, scan shift mode, low-voltage modes — each with its own fast and slow corners. A short path that passes hold in functional mode can fail in scan mode, where the clocking and data paths differ. Hold has to be clean in every mode.</p>
<p><strong>Second, on-chip variation moves the goalposts.</strong> In STA, a corner assumes the whole chip is at one process, voltage and temperature. Real silicon does not work like that: two identical cells on the same die switch at slightly different speeds because of process gradients, local supply droop and thermal differences. STA models this with on-chip variation (OCV) derating — for a hold check, the launch path is derated to be faster and the capture path slower, the pessimistic combination. Advanced forms (AOCV, which shrinks the derate for longer paths, and POCV, which treats variation statistically) reduce the pessimism, but the variation on a single die can still turn a path that passes at the nominal fast corner into a violation.</p>
<p><strong>Third, temperature inversion breaks the old rule entirely.</strong> For most of the history of CMOS, hot silicon was slow silicon. At low supply voltages that rule reverses: heat lowers the threshold voltage, which speeds the transistor up, and at a low enough supply that effect wins over the mobility loss heat causes. Cold silicon can then be the slowest and hot silicon the fastest. A design signed off only at the classic fast corner (cold) can miss a hold violation that only appears hot. At advanced nodes the honest answer is to check both temperature extremes for both setup and hold, and let the tool find the true worst case.</p>
<h2>3. Skew: the same knob that helps setup hurts hold</h2>
<p>Clock skew is the difference in clock arrival time between the launch flop and the capture flop. It is the one term a physical design engineer controls directly, through clock tree synthesis — and it is a double-edged one.</p>
<p>If the capture clock arrives <em>later</em> than the launch clock (positive skew), setup gets easier: the data has more time before the capture edge. But hold gets harder by exactly the same amount: the capture edge waits longer, giving the racing data more time to corrupt the capture. Negative skew does the reverse: it helps hold and hurts setup. There is no free lunch in the clock tree. It is also why hold violations often appear only <em>after</em> clock tree synthesis: before CTS the clock is ideal and skew is zero, so short paths pass hold trivially; once the tree gives the clock real arrival times, paths that were never a problem start violating hold. The same trade-off makes <em>useful skew</em> powerful and dangerous: deliberately delaying a capture clock lends time to a critical setup path, but every picosecond added to help setup is a picosecond taken from hold on the paths around it.</p>
<h2>4. Why hold is more dangerous than setup</h2>
<p>A setup violation is a speed problem. The data arrives too late for the target frequency, so the chip must run slower than planned. That is bad — it can mean missing a performance target — but it is a problem you can live with, work around, or fix by lowering the clock. In bring-up, a setup problem shows up as a maximum frequency limit, and the chip still functions below it.</p>
<p>A hold violation is a functional problem. The data arrives too early and corrupts the value being captured. The chip computes the wrong answer. And because the hold equation has no clock-period term, no frequency is safe: the violation is present at 10 MHz exactly as it is at 1 GHz. There is no workaround in the lab — no slower clock, no voltage tweak. If a hold violation escapes to silicon, the affected logic simply does not work, and the only remedies are a respin or a workaround that avoids exercising the failing paths, if one even exists.</p>
<p>That is why signoff treats the two checks differently. A setup violation with a few picoseconds of negative slack is a problem to close; a hold violation of the same size is a tapeout blocker. Hold must be zero, everywhere, in every corner and every mode, before the design ships.</p>
<h2>5. How hold actually gets fixed</h2>
<p>Since hold is a min-delay problem, every fix makes the short path slower. The tools and the engineer have a small set of levers, and each one costs something.</p>
<p><strong>Delay buffers and delay cells.</strong> The standard fix: insert one or more delay buffers on the short path, usually close to the capture flop. Each buffer adds a small, predictable amount of delay. The cost is real: every added buffer is area, leakage and dynamic power that the chip pays for the rest of its life, and a design with thousands of hold buffers pays a permanent tax for its short paths.</p>
<p><strong>Downsizing and higher-threshold cells.</strong> Replacing a fast, strong cell on the short path with a smaller drive-strength version, or swapping a low-threshold cell for a high-threshold one, slows the path down without adding cells. High-threshold cells are slower and leak less, so this fix can even help power — but it is a smaller lever than a buffer.</p>
<p><strong>Clock tree rebalancing.</strong> When a hold violation is large, the problem is often the clock, not the data. A big skew between launch and capture can make an otherwise reasonable path violate hold. Trimming the capture flop's clock branch removes the skew at its source. This is the fix to reach for when many violations cluster around the same flops, because it treats the cause instead of padding every path individually.</p>
<p>All of these fixes share one rule: fix hold at the corner where hold is worst — usually the fast corner — and then re-check setup at the slow corner. Every picosecond of delay added to fix hold eats into setup margin. A fix that clears hold at the fast corner and creates a setup violation at the slow corner is not a fix; it is a transfer of the problem from one corner to another. The two checks have to be closed together, in the same run.</p>
<h2>6. Closing hold: the ECO flow and the habits that matter</h2>
<p>Late in the flow, hold fixes are done as engineering change orders (ECOs): small, targeted changes to an already placed and routed design. The tool changes only the cells on the violating paths and re-routes just those nets, leaving the rest of the design untouched. A timing ECO resizes or buffers a handful of cells; a metal-only ECO, used after base-layer masks are made, works with pre-placed spare cells and changes only metal layers. Every ECO is followed by the same discipline: equivalence checking where logic changed, STA at all corners, and DRC and LVS. An ECO that is not re-verified is not a fix.</p>
<p>A few working habits keep hold from becoming a crisis at the end:</p>
<ul>
<li><strong>Fix hold after the clock tree exists, not before.</strong> Hold violations before CTS are mostly fiction — the ideal clock hides them. The real hold picture appears once CTS has given the clock real arrival times.</li>
<li><strong>Never sign off setup and hold from separate runs.</strong> They share paths, they share the clock tree, and they pull in opposite directions. A setup fix that delays a capture clock can create hold violations on neighbouring short paths; a hold fix that adds delay can break setup. Check them together, every time.</li>
<li><strong>Watch the short paths early.</strong> The paths that violate hold are the shortest ones — a flop feeding another flop with almost no logic between them, a scan path, a clock-gating control path. They are easy to ignore because their setup slack is enormous. Give them a minimum-delay budget the same way long paths get a maximum-delay budget.</li>
<li><strong>Treat a large hold violation as a clock problem first.</strong> A handful of picoseconds on one path is a buffer. Tens of picoseconds across many paths sharing a flop is skew, and padding every path is the wrong answer.</li>
<li><strong>Respect the margin, but do not gold-plate it.</strong> Hold fixing adds cells that burn power forever. Fix to the signoff requirement with the required margin, and no further.</li>
</ul>
<p>Hold time is the quiet constraint in timing signoff. It has no frequency to tune, no clock to slow, and no second chance after tapeout. It fails at the fast corner, in the mode you forgot, under variation, and at the temperature you assumed was safe — so check it everywhere, fix it structurally, and never let a setup fix walk out the door without a hold check beside it.</p>
<h2>Sources</h2>
<ul>
<li>SiliconPath — Setup Time, Hold Time &amp; Slack: Static Timing Analysis — <a href="https://jobs.siliconpath.in/learn/static-timing-analysis/setup-hold-slack-analysis">https://jobs.siliconpath.in/learn/static-timing-analysis/setup-hold-slack-analysis</a></li>
<li>ChipVerify — Setup and Hold Checks — <a href="https://chipverify.com/static-timing-analysis/setup-and-hold-checks">https://chipverify.com/static-timing-analysis/setup-and-hold-checks</a></li>
<li>ChipBuddy — Setup and Hold Checks — <a href="https://www.chipbuddyvlsi.com/vol10/5/">https://www.chipbuddyvlsi.com/vol10/5/</a></li>
<li>VLSI Web — STA Interview Questions: Setup, Hold, SDC and Sign-off Explained — <a href="https://vlsiweb.com/sta-interview-questions/">https://vlsiweb.com/sta-interview-questions/</a></li>
<li>SiliconPath — Timing Closure: Fixing Setup &amp; Hold Violations — <a href="https://jobs.siliconpath.in/learn/static-timing-analysis/timing-closure-techniques">https://jobs.siliconpath.in/learn/static-timing-analysis/timing-closure-techniques</a></li>
<li>VLSI Web — Timing Closure in Physical Design: Fixing Setup, Hold and ECOs — <a href="https://vlsiweb.com/sta-in-pd/">https://vlsiweb.com/sta-in-pd/</a></li>
<li>VLSIDesignFlow — Hold Slack &amp; Minimum Delay Analysis — <a href="https://www.vlsidesignflow.com/hold-slack/">https://www.vlsidesignflow.com/hold-slack/</a></li>
<li>iVLSI — Temperature Inversion in VLSI: Causes, STA Corners &amp; Fixes — <a href="https://data.ivlsi.com/temperature-inversion-in-vlsi/">https://data.ivlsi.com/temperature-inversion-in-vlsi/</a></li>
<li>ChipVerify — Temperature Inversion: Static Timing Analysis — <a href="https://chipverify.com/static-timing-analysis/temperature-inversion">https://chipverify.com/static-timing-analysis/temperature-inversion</a></li>
<li>ChipBuddy — Operating Conditions, PVT Corners &amp; Derating — <a href="https://www.chipbuddyvlsi.com/vol10/9/">https://www.chipbuddyvlsi.com/vol10/9/</a></li>
<li>iVLSI — Static Timing Analysis (STA) in VLSI: Setup, Hold, OCV/AOCV/POCV, CRPR — <a href="https://data.ivlsi.com/static-timing-analysis-sta-in-vlsi/">https://data.ivlsi.com/static-timing-analysis-sta-in-vlsi/</a></li>
<li>VLSI Verification — Static Timing Analysis: Setup Time, Hold Time &amp; Slack Calculation — <a href="https://www.vlsiverification.net/tutorials/digital/sta.html">https://www.vlsiverification.net/tutorials/digital/sta.html</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[The Wafer Got Bigger: Working Power Devices on 300 mm Silicon Carbide — This Week in Semiconductors]]></title><description><![CDATA[Last week the industry's story was that memory is sold out for years. This week the story moved from the memory array to the power socket. On November 9…]]></description><link>https://thetapeoutlog.hashnode.dev/the-wafer-got-bigger-working-power-devices-on-300-mm-silicon-carbide-this-week-in-semiconductors</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/the-wafer-got-bigger-working-power-devices-on-300-mm-silicon-carbide-this-week-in-semiconductors</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Fri, 13 Nov 2026 04:01:10 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/6622c9e9-e216-405f-a4a4-ce6d521f4fca.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><img src="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/6622c9e9-e216-405f-a4a4-ce6d521f4fca.webp" alt="" /></p>
<p>Last week the industry's story was that memory is sold out for years. This week the story moved from the memory array to the power socket. On November 9, Infineon announced it had processed 300-millimeter silicon carbide wafers into working power devices inside a full-scale fab — and said it is the first company to do so. In the same week, a competitor walked away from a silicon carbide expansion worth up to $2 billion. One milestone, one retreat, and the same material at the center of both.</p>
<h2>What actually happened</h2>
<p>Infineon said it turned 300 mm silicon carbide (SiC) wafers into functional power semiconductor devices in a production fab environment, using manufacturing equipment already in place for silicon. The company describes the result as an industry first. A 300 mm wafer yields 2.3 times as many chips as a 200 mm wafer, by the company's count. The first processed wafer went on public display at the electronica trade show, running November 10 to 13.</p>
<p>Two details matter more than the headline number. The wafers were processed on existing silicon tools. And Infineon also demonstrated that its Cold Split wafering process — which separates wafers from a crystal with less material loss and much shorter processing time than conventional wire sawing — works on 300 mm material. Both point the same way: this was a manufacturing result, not a laboratory demonstration.</p>
<p>It is worth being precise about what was shown. Infineon has working devices on 300 mm SiC wafers. It has not given a production date, yields, cost figures, device ratings, or a named customer. It will plan the ramp timing with customers and partners, while continuing its 200 mm SiC roadmap. This is process feasibility, not volume production.</p>
<h2>Why wafer size is the whole game</h2>
<p>A fab processes one wafer at a time. The more chips that come off each wafer, the lower the cost of each chip — and the arithmetic is generous. Wafer area scales with the square of the diameter, so moving from 200 mm to 300 mm multiplies the usable area by 2.25. Infineon's 2.3x chips-per-wafer figure lines up with that geometry. For a PD engineer, the lesson is familiar: the wafer is the unit of cost, and everything downstream divides by the number of good dies on it.</p>
<p>That is why the industry has chased larger wafers for decades, and why silicon carbide has lagged so far behind. Silicon moved to 300 mm more than two decades ago, while most of the SiC industry is still moving from 150 mm to 200 mm. The material is nearly as hard as diamond, crystals grow slowly at extreme temperatures, and slicing them wastes expensive material. Getting a 300 mm SiC wafer is hard; getting working devices out of one, on standard tools, is the step that changes the economics.</p>
<h2>Why silicon carbide, and why now</h2>
<p>Silicon carbide switches high voltages with less energy lost as heat than silicon, which is why it has become the material of choice for traction inverters, on-board chargers, fast chargers, large-scale energy storage, industrial drives, and uninterruptible power supplies. Demand is also rising in AI data centers, where SiC is finding roles in solid-state transformers, circuit breakers, and power supply units.</p>
<p>Power delivery has become one of the binding constraints in AI infrastructure, and every watt lost in conversion is a watt that has to be cooled. A cheaper SiC die does not solve that by itself — the die is small, runs hot, and switches fast, so heat and stray inductance in the package limit what it can do — but it lowers the cost of the most expensive step in the chain.</p>
<h2>The other headline of the same week</h2>
<p>Three days before Infineon's announcement, onsemi and a government ministry jointly confirmed they will not proceed with a planned silicon carbide expansion first announced in June 2024, a project valued at up to $2 billion. The decision was attributed to changing conditions in the SiC market and the project's economics. Existing silicon and SiC production at the site continues unchanged.</p>
<p>The contrast is the story. One company is proving that 300 mm SiC can be built on existing tools; another has decided that new SiC capacity does not pay at current prices. Both can be right at the same time. Electric vehicle sales have grown more slowly than chipmakers expected when they announced their new fabs, and wafer prices have fallen. A wafer that yields 2.3 times as many chips arrives just when there are questions about whether there are too many chips.</p>
<p>Keep one distinction straight. In January, Wolfspeed said it had produced a single-crystal 300 mm silicon carbide wafer — the material. Infineon has now shown working devices built on that size of wafer. A wafer is not a finished device, and neither company has shown volume production — but the sequence matters: material first, devices second, volume third.</p>
<h2>What to watch next</h2>
<p>Four things will show which way this goes. First, whether Infineon names a start date or a first customer for 300 mm SiC. Second, what happens to 200 mm SiC wafer prices through 2027 — cheaper wafers could expand the market, or squeeze everyone who invested assuming old prices would last. Third, whether more SiC factory projects are cancelled or delayed. And fourth, whether carmakers qualify 300 mm SiC devices for traction inverters, a process that takes years of heat, vibration, and voltage stress testing and will not be rushed on a supplier's schedule.</p>
<p>For now, the practical read is simple. The 200 mm roadmap will carry most of the volume for years, and decisions about substrates, sintering, and module packaging define heat and reliability today. But the wafer got bigger this week, on standard tools, in a real fab — and the harder work of making it pay has only started.</p>
<h2>Sources</h2>
<ul>
<li>Infineon Technologies, "Infineon marks innovation milestone with world's first 300-millimeter power silicon carbide wafer," November 9, 2026 — <a href="https://www.infineon.com/press-release/2026/infpr202611-028">https://www.infineon.com/press-release/2026/infpr202611-028</a></li>
<li>Sparknify, "Infineon's 300 mm Silicon Carbide Wafer: What Changes," November 10, 2026 — <a href="https://www.sparknify.com/post/20261110-the-wafer-got-bigger-en">https://www.sparknify.com/post/20261110-the-wafer-got-bigger-en</a></li>
<li>Interesting Engineering, "New 300-mm silicon carbide wafers could boost production of EV chips," November 11, 2026 — <a href="https://site-test.interestingengineering.com/innovation/300-mm-silicon-carbide-wafers-built">https://site-test.interestingengineering.com/innovation/300-mm-silicon-carbide-wafers-built</a></li>
<li>VoltGrid News, "Infineon Shows 300 mm SiC Devices, Claims 2.3x Chips per Wafer," November 9, 2026 — <a href="https://voltgridnews.com/story/infineon-shows-300-mm-sic-devices-claims-2-3x-chips-per-wafer">https://voltgridnews.com/story/infineon-shows-300-mm-sic-devices-claims-2-3x-chips-per-wafer</a></li>
<li>The Code W, "Semiconductor Watch," November 9, 2026 — <a href="https://www.thecodew.com/2026/11/semiconductor-watch-november-9.html">https://www.thecodew.com/2026/11/semiconductor-watch-november-9.html</a></li>
<li>De Vector, "Onsemi scraps planned chip expansion," November 6, 2026 — <a href="https://www.devector.nl/en/edition/2026-11-06/onsemi-annuleert-chipuitbreiding.html">https://www.devector.nl/en/edition/2026-11-06/onsemi-annuleert-chipuitbreiding.html</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[PD for AI Accelerators: What Changes When the Chip Is One Giant MAC Array]]></title><description><![CDATA[By Pradeep Kumar Adusumalli Most chips are a city: a little of everything, stitched together. An AI accelerator is one enormous factory floor — a single…]]></description><link>https://thetapeoutlog.hashnode.dev/pd-for-ai-accelerators-what-changes-when-the-chip-is-one-giant-mac-array</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/pd-for-ai-accelerators-what-changes-when-the-chip-is-one-giant-mac-array</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Thu, 12 Nov 2026 10:59:37 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/d40203f8-fc76-4371-9075-61f0a7f4b3c7.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>By Pradeep Kumar Adusumalli</em></p>
<p>Most chips are a city: a little of everything, stitched together. An AI accelerator is one enormous factory floor — a single multiply-accumulate (MAC) array, tiled thousands of times over, with everything else pushed to the edges — and that changes everything downstream: floorplan, power grid, clock tree, thermal plan, signoff.</p>
<p>The array is an old idea. Systolic arrays, proposed by Kung and Leiserson in 1978, are grids of processing elements where each talks only to its nearest neighbours, data advances one hop per clock tick, and each value is reused by many elements before leaving the array. Modern AI hardware revived them: in a weight-stationary mapping, each weight is loaded once and stays resident while inputs stream past, so one fetch feeds many operations. The PD problem, then, is not "how do I place a complex SoC" — it is "how do I build, power, clock and cool one giant regular structure without letting its scale defeat me".</p>
<h2>1. The floorplan stops being a floorplan</h2>
<p>In a normal SoC, floorplanning is negotiation between differently shaped blocks; in an accelerator, the array <em>is</em> the floorplan. You design one processing element, one tile, then tile it — and the tile's quality decides the chip's. The reason is wire delay: the scratchpad memories feeding the array must sit close to its edges, because every millimetre there is wire delay on the critical data path. Published work on an open systolic-array generator states this directly: large arrays significantly impact physical design and cycle time, since the scratchpad memories need placing to reduce wire delay between memory and array edges. Those designs taped out in 16 nm and 22 nm-class processes and met timing and power — because memory placement was a first-class floorplan decision.</p>
<p>Make the tile self-contained — own power straps, own clock entry, clean abutment edges — so tiling it hundreds of times creates no new problem at any seam. Keep the memories at the array edge with short, wide data paths, and plan the array edge as a real interface: weight loading, streaming, partial-sum drain and control all cross it, so give it dedicated channels early. Size the array to the workload, not the die: large arrays are slow and underused on small matrices, while an idle array still burns leakage and needs its clock tree and power grid.</p>
<h2>2. Power delivery: the grid becomes the floorplan</h2>
<p>This is where accelerator scale bites hardest. Modern AI accelerators draw currents exceeding 1,000 A at 0.6 to 0.85 V core, with total board power of 700 W to over 1,000 W per card — a sustained 1,250 A at 0.8 V. And it is not steady: workloads switch from idle to fully loaded in an instant, with transient slew rates above 1,000 A per microsecond.</p>
<p>A 20 mV IR drop on a 0.8 V rail is a 2.5% loss of supply voltage, enough to cause clock errors and chip resets. Holding voltage within ±3% at 1,000 A needs target impedance below 0.1 mΩ — an on-die grid problem, planned with the floorplan. Current rises, IR drop rises, local supply falls, cell delay rises, timing margin falls — so a path that closes under ideal voltage can fail where dynamic droop bites. And the droop is not random: the centre of a large active array is the worst place on the die, furthest from supply entry and surrounded by switching neighbours.</p>
<p>Budget IR drop per region and run dynamic IR analysis with realistic switching activity — a vectorless average hides the droop that kills you. Spend metal on power early: a stronger grid improves IR drop and electromigration, but wider power metal consumes routing resources, so fix the grid pitch before signal routing starts. Backside power delivery — a dedicated plane beneath the transistor layer — frees front-side metal for signals. Place decoupling where the current switches, staged from bulk to high-frequency with the shortest loop; a decap three millimetres away is decoration. Model the transient, not just the DC drop: a grid passing static IR can still fail when the whole array wakes in the same cycle.</p>
<h2>3. Clocking a million-flop array</h2>
<p>A large tensor engine can contain millions of sequential elements, and its clock network must control skew, insertion delay, on-chip-variation sensitivity, transition quality, clock power and routing congestion all at once. The systolic structure makes it sharper: data moves as a wavefront, and each element must receive its input in the same cycle as the partial sum from its neighbour above, or its sum is wrong. In the original 1978 analysis, large two-dimensional arrays could require slowing the global clock to compensate for skew.</p>
<p>Treat the array as one clock domain with one skew budget: a mesh or balanced H-tree across the array, built as a single structure, beats per-tile optimisation — do not let the tree fragment into islands with different insertion delays. Budget skew against the wavefront, not just setup and hold: local skew between neighbours matters more than global skew across the die. Model on-chip variation across the array, not at its corners: a regular array spans the full die, so a uniform-variation corner is optimistic. Count clock power as a first-class number — dynamic power scales with activity, capacitance, voltage squared and frequency, and activity in a tensor array is very high. Start multi-mode, multi-corner analysis early: clock, power, thermal and variation effects move together, and late MMMC surprises in an array this regular repeat at every tile.</p>
<h2>4. Heat: the hotspot is the spec</h2>
<p>The maximum frequency and power of today's advanced chips is set not by average die temperature but by its hottest local region. Designs may target a 105 °C hotspot against an 85 °C average; a hotspot 50 to 100 µm across can constrain the clock of a much larger block, and 30 °C differences can exist across a few hundred microns. The physics is a feedback loop: activity, power, temperature and leakage rise together, while heat slows critical paths. The lifetime cost is steep — every 10 °C of sustained junction temperature can roughly halve lifetime for a 0.7 eV failure mechanism (one published example: 1,000 hours at 125 °C ≈ nine years at 55 °C).</p>
<p>What makes accelerators special is that the hotspot moves with the workload. Recent open-access research on systolic arrays found sparse workloads consistently produce the largest thermal spread, while dense workloads stay more uniform: only part of the array is active, so the active region heats while its neighbours stay cool — and the gradient, not the average, stresses timing. Correctness risk, the study found, depends on workload-driven nonuniformity, dataflow organisation and exposed path classes — a hotspot can stay benign, be masked, or propagate into silent corruption.</p>
<p>The memory sets a harder limit than the logic: the MAC arithmetic would happily keep operating, but the nearby SRAM determines the allowable voltage, frequency and temperature operating point, and DRAM bonded into the stack must refresh twice as often above 85 °C. In one published study, HBM stacked on a GPU drove it to 141.7 °C, and the fix included halving the clock at a 28% performance cost. Never put memory on top of hot logic.</p>
<p>Do thermal-aware placement, not just thermal analysis: spread high-activity blocks, keep the hottest logic away from the SRAM that sets its operating point, and treat the thermal map as a floorplan input. Sign off against the worst workload, not the average — dense activity is the easy case; sparse, bursty activity creates the steepest gradients. Close timing at hotspot temperature, iterate the leakage loop until power and temperature converge, and use throttling and DVFS as a designed feature: a planned throttle point protects the chip; an unplanned one is a product failure.</p>
<h2>5. Routing congestion in and around the array</h2>
<p>Inside the array, routing is almost free: nearest-neighbour connections are short and local, which is the whole point of the systolic structure. The congestion lives at the boundaries — where data enters, where partial sums leave, where the memory interfaces sit, and where control threads through. The memory edge is the worst of it. High-bandwidth memory is DRAM dies stacked vertically, joined by through-silicon vias, mounted next to the processor on a silicon interposer: a 2,048-bit interface, up to 8 Gb/s per pin and up to 2 TB/s per stack, with stacks within a millimetre of the processor. All that bandwidth must escape through the array's periphery — thousands of signals where the power grid is densest and the clock tree enters. The package matters too: micro-bump pitches below 55 µm, packages up to 100 mm × 100 mm, and chiplet core-voltage tolerances often ±2% or less push power integrity back onto the die-level grid.</p>
<p>Keep the memory PHYs close with wide channels — short, wide paths beat long, narrow ones — and place the PHYs before the array tiles, since they consume the edge the array wants. Reserve routing channels for the wide buses early: routed after the array is placed, they will be routed through it, and through its power grid. Use higher metals for long runs, and run congestion analysis before the floorplan is frozen — congestion appears all at once at the array edge, late, when fixing it means moving the array. Power and signal metal compete for the same layers, so decide the grid pitch and the bus plan together.</p>
<h2>6. Signoff corners that matter most at this scale</h2>
<p>Signoff for an accelerator is not a longer version of SoC signoff — the failure modes are different.</p>
<ul>
<li><strong>Dynamic IR, not just static:</strong> the killer corner is a workload step — the whole array waking in the same cycle, at low voltage, at high temperature. Static IR analysis will not find it.</li>
<li><strong>Thermal corners driven by workload, not ambient:</strong> sign off at the hotspot temperature the worst workload creates. A uniform-temperature corner is a fiction at this power density.</li>
<li><strong>Voltage-droop-aware timing:</strong> a path closing under ideal voltage may fail where dynamic droop bites. Signoff must see the droop map, not just the nominal supply.</li>
<li><strong>Sparse-workload corners:</strong> a suite exercising only dense, uniform activity misses the workload with the largest thermal spread and steepest gradients.</li>
<li><strong>Ageing corners:</strong> persistent local thermal stress interacts with negative-bias temperature instability and other ageing mechanisms — hotspots are a lifetime reliability problem, not just a timing problem.</li>
<li><strong>Multi-mode, multi-corner from the start:</strong> clock, power, thermal and variation effects interact. Analysing them separately is how a design passes every check and fails in silicon.</li>
</ul>
<h2>Closing</h2>
<p>The MAC array is a gift to physical design: regularity, locality, reuse, a structure that tiles cleanly. But scale turns every virtue into a demand — a problem in one tile is a problem in every tile, the routing problem moves to the edges, and switching together means current spikes, voltage droop and hotspots that move with the workload. Design around what the array does at scale: memories at its edge, power fed as if the grid were the floorplan, one clock structure, signoff at the temperature the worst workload creates — and never let the average hide the hotspot.</p>
<h2>Sources</h2>
<ul>
<li>Gemmini: an open systolic-array generator — array tiling, scratchpad placement and wire delay, tapeouts in 16 nm and 22 nm-class processes: <a href="https://arxiv.org/html/1911.09925v2">https://arxiv.org/html/1911.09925v2</a></li>
<li>Physical design challenges of LLM accelerators — IR-drop chain, thermal feedback loop, clock-network demands, congestion and MMMC: <a href="https://www.guriai.com/blog/llm-ai-chip-physical-design.html">https://www.guriai.com/blog/llm-ai-chip-physical-design.html</a></li>
<li>Hotspots and their effect on frequency, lifetime and reliability — hotspot vs average temperature, 10 °C lifetime rule, SRAM as the operating-point constraint: <a href="https://www.df.com/insights/hotspots-are-evil">https://www.df.com/insights/hotspots-are-evil</a></li>
<li>High-bandwidth memory explained — HBM4 standard, 2,048-bit interface, 2 TB/s per stack, interposer packaging: <a href="https://www.signalstack.news/blog/what-is-hbm-high-bandwidth-memory/">https://www.signalstack.news/blog/what-is-hbm-high-bandwidth-memory/</a></li>
<li>AI accelerator power delivery network design — 1,000 A+ currents, 0.6–0.85 V core, target impedance below 0.1 mΩ, transient slew above 1,000 A/µs: <a href="https://www.nextpcb.com/blog/ai-accelerator-pcb-pdn-design">https://www.nextpcb.com/blog/ai-accelerator-pcb-pdn-design</a></li>
<li>Systolic arrays from 1978 to modern AI matrix engines — Kung–Leiserson origins, wavefront timing, clock-skew limits on large arrays: <a href="https://hsmahesha.github.io/papers/systolic-arrays.pdf">https://hsmahesha.github.io/papers/systolic-arrays.pdf</a></li>
<li>On-package power delivery for AI accelerators — 48 V to 0.8 V conversion, backside power delivery, current density and IR budgets: <a href="https://adikumar.co/the-on-package-delivery-stack/">https://adikumar.co/the-on-package-delivery-stack/</a></li>
<li>Thermal nonuniformity-aware reliability screening for systolic AI accelerators — sparse workloads and thermal spread, thermal-to-timing stress, silent corruption: <a href="https://mdpi-res.com/d_attachment/jlpea/jlpea-16-00018/article_deploy/jlpea-16-00018.pdf?version=1780230528">https://mdpi-res.com/d_attachment/jlpea/jlpea-16-00018/article_deploy/jlpea-16-00018.pdf?version=1780230528</a></li>
<li>Stacking, memory and heat — HBM-on-GPU thermal study, DRAM temperature limits, memory placement: <a href="https://www.evaporion.com/research/stacking-memory-heat/">https://www.evaporion.com/research/stacking-memory-heat/</a></li>
<li>Chiplet architecture, CoWoS and EMIB — interposer packaging, micro-bump pitch, ±2% voltage tolerance, large-package mechanical stress: <a href="https://morepcb.com/chiplet-architecture-cowos-emib-pcb-design-2026/">https://morepcb.com/chiplet-architecture-cowos-emib-pcb-design-2026/</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[What Interviewers Actually Ask in a PD Interview]]></title><description><![CDATA[By Pradeep Kumar Adusumalli A physical design interview is not a quiz on definitions. The interviewer already knows you can recite what CTS stands for…]]></description><link>https://thetapeoutlog.hashnode.dev/what-interviewers-actually-ask-in-a-pd-interview</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/what-interviewers-actually-ask-in-a-pd-interview</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Wed, 11 Nov 2026 04:05:48 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/7e7245b5-ed76-4b1d-b6e3-48765e72129c.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>By Pradeep Kumar Adusumalli</em></p>
<p>A physical design interview is not a quiz on definitions. The interviewer already knows you can recite what CTS stands for. What they are really testing is whether you can take a netlist to a clean, sign-off-ready layout — and whether you understand <em>why</em> each step exists. Below are the questions that come up again and again, and what a strong answer shows.</p>
<h2>The opener: walk me through the flow</h2>
<p><strong>"Walk me through the RTL-to-GDSII flow."</strong></p>
<p>The canonical first question. A strong answer names the stages in order — synthesis, floorplanning, power planning, placement, clock tree synthesis, routing, and signoff — and for each stage states its goal and its main checks. What separates seniors: they also say what can go wrong at each stage. Expect the follow-up: "timing fails after routing — what do you check first?"</p>
<h2>Floorplanning</h2>
<p><strong>"What makes a good floorplan?"</strong></p>
<p>One that minimises wire length and congestion, places macros sensibly, keeps related logic together, and leaves room for the clock tree and the power grid. A bad floorplan cannot be rescued downstream.</p>
<p><strong>"What is utilisation, and what target do you aim for?"</strong></p>
<p>Utilisation is standard-cell area divided by available core area. A typical starting target quoted in interview preparation material is around 60–75%. Too high, and you get congestion and no room for buffers; too low, and area is wasted. Memory-heavy designs often start lower, around 40–60%.</p>
<p><strong>"How do you place macros?"</strong></p>
<p>Toward the edges or corners, so they do not block routing in the centre, with a halo around each one (guides commonly quote 2–5 µm), pins facing the routing channels, and related macros grouped together. Long flylines across the chip after floorplanning are a preview of congestion.</p>
<h2>Placement</h2>
<p><strong>"What does placement optimise, and what is congestion?"</strong></p>
<p>Placement positions every standard cell to minimise timing-critical wire length and congestion while keeping the placement legal — cells on rows, on the power-rail grid, without overlaps. Congestion is when a region demands more routing than the available tracks can supply. Fix it by spreading cells or adjusting the floorplan.</p>
<p><strong>"What is timing-driven placement?"</strong></p>
<p>A pure wire-length minimiser might spread the cells of a critical path apart. Timing-driven placement pulls the cells on critical paths close together so wire delay stays small, and lets non-critical paths tolerate longer wires.</p>
<h2>Clock tree synthesis</h2>
<p><strong>"What is CTS, and what are skew and insertion delay?"</strong></p>
<p>CTS builds the network that distributes the clock to every sequential element. Skew is the difference in clock arrival time between two flip-flops; insertion delay is the delay from the clock source to the flops. The goal is to balance skew while controlling latency and clock power.</p>
<p><strong>"What is useful skew?"</strong></p>
<p>Deliberately delaying or advancing some clock arrivals to borrow time across a critical path: a capture flop whose clock arrives later relaxes a tight setup path, at the cost of hold margin. It separates engineers who think of skew as "always bad" from engineers who treat it as a closure tool. Positive skew helps setup and hurts hold; negative skew does the opposite. How much is acceptable? No single number — skew is judged by slack, not by a magic number.</p>
<h2>Routing</h2>
<p><strong>"What is the difference between global and detailed routing?"</strong></p>
<p>Global routing divides the die into a coarse grid and assigns each net to a sequence of regions and layers, estimating congestion without producing real wire geometry. Detailed routing then creates the exact wires, vias, and layer assignments that satisfy every design rule.</p>
<p><strong>"What is crosstalk, and how do you fix it?"</strong></p>
<p>A switching wire (the aggressor) couples noise onto a neighbouring wire (the victim) through capacitance. It changes delay — same-direction switching speeds the victim up, opposite-direction slows it down — and it can inject a glitch on a quiet net. Fixes: shield critical nets with supply or ground, increase spacing, or use lower metal layers.</p>
<h2>Timing closure and STA</h2>
<p><strong>"What are setup and hold violations, and how do you fix each?"</strong></p>
<p>A setup violation means the data path is too slow for the clock period; fix it by making the path faster — upsizing cells, restructuring logic, or using useful skew. A hold violation means data races through too fast; fix it by adding delay buffers. Setup is fixed by making the data path faster, hold by making it slower.</p>
<p><strong>"Why is setup closed at the slow corner and hold at the fast corner?"</strong></p>
<p>Setup is worst when cells are slowest — slow process, low voltage, high temperature — so you close setup at the slow corner. Hold is worst when cells are fastest, so you fix hold at the fast corner. Mixing these two up is an instant red flag.</p>
<p><strong>"What is STA, and how is it different from dynamic timing simulation?"</strong></p>
<p>Static timing analysis checks every timing path against the constraints without test stimulus — exhaustive and vectorless, which is why it scales to millions of paths. Dynamic simulation only checks the paths your vectors exercise. Signoff relies primarily on STA. Slack is required time minus arrival time; negative slack is a violation.</p>
<p><strong>"What are false paths and multicycle paths?"</strong></p>
<p>A false path is never functionally exercised — across an asynchronous boundary, or through mutually exclusive settings — and you declare it so STA does not block closure on it. A multicycle path is intentionally allowed more than one clock cycle. The trap: a wrong exception can mask a real violation and produce silicon that fails.</p>
<h2>DRC, LVS, and physical verification</h2>
<p><strong>"What checks make up physical verification?"</strong></p>
<p>DRC, LVS, antenna checks, and ERC. These gate tape-out: a single unresolved DRC or LVS mismatch blocks the chip. For DRC, be ready with three examples: spacing, width, and via enclosure violations.</p>
<p><strong>"What is LVS, and what errors does it catch?"</strong></p>
<p>LVS extracts a netlist from the physical layout and compares it to the reference netlist. It catches opens, shorts, and extra or missing devices. LVS must be clean before tape-out.</p>
<h2>Power: IR drop and electromigration</h2>
<p><strong>"What is IR drop, and how do you reduce it?"</strong></p>
<p>The voltage lost across the resistance of the power delivery network when current flows — V = I × R. Cells that see a lower supply switch more slowly, which can cause setup violations; severe IR drop can stop cells switching at all. Reduce it with wider power straps, more vias, and decoupling capacitors near the hotspots. Dynamic IR drop — the transient droop from simultaneous switching — is usually the harder problem.</p>
<p><strong>"What is electromigration, and how do you fix it?"</strong></p>
<p>The slow displacement of metal atoms in a wire under high current density, which over time creates voids (opens) or hillocks (shorts) and fails the chip. Fix it by widening the wire, adding parallel wires or more vias, and spreading high-current cells apart. Both are signoff checks, verified with dedicated power-integrity analysis before tape-out.</p>
<h2>Low power: UPF and power domains</h2>
<p><strong>"What is a power intent file (UPF)?"</strong></p>
<p>UPF — Unified Power Format, standardised as IEEE 1801 — describes the power architecture of a multi-voltage design: the power domains, their supply voltages, the power states, and the level shifters, isolation cells, and power switches needed at domain boundaries. The tool reads UPF and inserts those cells automatically.</p>
<p><strong>"What happens when a signal crosses between power domains?"</strong></p>
<p>It needs a level shifter to translate the signal from one voltage to the other. If a domain is powered off, isolation cells stop its outputs floating and corrupting the powered-on logic. Know the usual techniques — power gating, clock gating, multi-threshold cells, voltage islands — and what each saves and costs: power gating saves leakage but adds wake-up latency; clock gating saves dynamic power but needs care in the clock tree.</p>
<h2>Signoff: what must pass before tape-out</h2>
<p><strong>"What happens during signoff?"</strong></p>
<p>The final verification phase before the design is released to the foundry, and every check must pass: timing clean across all corners and modes with extracted parasitics, DRC and LVS clean, IR drop and electromigration within limits, antenna and ERC clean.</p>
<p><strong>"Why is SPEF needed for signoff timing?"</strong></p>
<p>SPEF carries the extracted RC parasitics of every wire in the physical layout. Wire delay depends on real metal resistance and capacitance, which are only known after layout. Pre-route timing uses estimated wire loads, which one interview guide notes can be off by 20–30%. Signoff STA runs on SPEF, which is why post-route timing can move even when the netlist has not changed.</p>
<h2>How to prepare using your own project work</h2>
<p>This is the part most candidates underprepare. Interviewers will spend more time on your own projects than on textbook questions.</p>
<p>Prepare three stories from your own work, each with the same shape: the situation (the block, the target frequency, the constraint that mattered); the symptom (the number you saw — worst negative slack, a congestion hotspot, an IR drop number); what you actually did (the lever you pulled, and why that one); and the result (the number after the fix). Keep a small log while you work — the violation, the root cause, the fix, the before-and-after numbers. That log beats any list of memorised answers, because every number in it is yours.</p>
<h2>Common mistakes</h2>
<ul>
<li><strong>Reciting definitions without the "why".</strong> Knowing CTS balances skew is table stakes. Knowing when to accept skew to close setup is the differentiator.</li>
<li><strong>Mixing up the corners.</strong> Setup at the slow corner, hold at the fast corner. Getting this backwards ends interviews.</li>
<li><strong>Treating signoff as someone else's job.</strong> Every answer should tie back to timing, DRC, LVS, or power closure.</li>
<li><strong>Claiming tool expertise you cannot demonstrate.</strong> Be honest about what you have actually run.</li>
<li><strong>No numbers.</strong> "I fixed timing" means nothing. "I closed a block with 120 ps of worst negative slack by restructuring a path and moving two macros" means everything.</li>
</ul>
<h2>Closing</h2>
<p>A PD interview is really one question asked many ways: can you take a design from a netlist to working silicon? The engineers who do well are not the ones with the most memorised answers. They are the ones who can look at a violation, trace it to its root cause across the flow, and explain the fix in terms of the physics and the numbers. Prepare that way, and the interview starts feeling like a design review.</p>
<h2>Sources</h2>
<ul>
<li>MockVise — VLSI Physical Design Interview Questions (2026 Guide): <a href="https://www.mockvise.com/blog/vlsi-physical-design-interview-questions">https://www.mockvise.com/blog/vlsi-physical-design-interview-questions</a></li>
<li>MockVise — Static Timing Analysis (STA) Interview Questions &amp; Answers: <a href="https://www.mockvise.com/blog/static-timing-analysis-interview-questions">https://www.mockvise.com/blog/static-timing-analysis-interview-questions</a></li>
<li>SiliconPath — Physical Design Interview Questions (Top 30): <a href="https://jobs.siliconpath.in/learn/interview/physical-design-interview-questions-top-30">https://jobs.siliconpath.in/learn/interview/physical-design-interview-questions-top-30</a></li>
<li>Digital Electronics Blog — Top 50 Physical Design Interview Questions with Answers: <a href="http://blog.digitalelectronics.co.in/2025/01/top-50-physical-design-interview.html">http://blog.digitalelectronics.co.in/2025/01/top-50-physical-design-interview.html</a></li>
<li>Hirist — Top 25+ Physical Design Interview Questions (2026): <a href="https://www.hirist.tech/blog/top-25-physical-design-interview-questions-and-answers/">https://www.hirist.tech/blog/top-25-physical-design-interview-questions-and-answers/</a></li>
<li>SiliconVLSI — Physical Design Interview Questions for VLSI Engineers: <a href="https://siliconvlsi.com/physical-design-interview-questions/">https://siliconvlsi.com/physical-design-interview-questions/</a></li>
<li>LearnElectronicsIndia — Top 30 Interview Questions &amp; Answers for Physical Design Engineer: <a href="https://www.learnelectronicsindia.com/post/top-30-interview-questions-answers-for-physical-design-engineer">https://www.learnelectronicsindia.com/post/top-30-interview-questions-answers-for-physical-design-engineer</a></li>
<li>InterviewGemini — Interview Questions for Physical Verification: <a href="https://skills.interviewgemini.com/interviews/interview-questions-for-physical-verification/">https://skills.interviewgemini.com/interviews/interview-questions-for-physical-verification/</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[ECO Flows: Fixing Silicon Bugs Without a Full Re-Route]]></title><description><![CDATA[By Pradeep Kumar Adusumalli Every physical design engineer knows the moment. The design is placed, routed, and nearly signed off — and then verification…]]></description><link>https://thetapeoutlog.hashnode.dev/eco-flows-fixing-silicon-bugs-without-a-full-re-route</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/eco-flows-fixing-silicon-bugs-without-a-full-re-route</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Tue, 10 Nov 2026 04:07:06 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/8aa9e190-b4d1-423c-8732-c19338bf4366.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>By Pradeep Kumar Adusumalli</em></p>
<p>Every physical design engineer knows the moment. The design is placed, routed, and nearly signed off — and then verification finds a bug, or the last STA run shows a handful of violating paths. Re-running the whole flow would take days, disturb everything already working, and probably create new violations elsewhere. So instead, you do an ECO.</p>
<p>An engineering change order is the chip industry's version of keyhole surgery: a small, targeted change to a design that is already implemented, made without re-running synthesis and place-and-route from scratch. This is how it actually works.</p>
<h2>What an ECO is — and why it happens</h2>
<p>An ECO is a controlled edit to a netlist and layout that is already placed and routed: instead of going back to synthesis, you change only the cells and wires that must change and leave the rest exactly as it is. By ECO stage, a design carries weeks of converged placement, clock-tree work and routing; a fresh run would move cells, rebuild the clock tree and re-route everything, and every result already checked would have to be checked again.</p>
<p>ECOs come in three kinds, and telling them apart matters because each has different limits.</p>
<p>A <strong>functional ECO</strong> changes the logic itself — a bug found in late verification, a spec change, a wrong polarity on a control signal. The RTL is fixed and the netlist is patched to match, and the patched netlist must be proven equivalent to the new RTL: a formal verification job, not a simulation one.</p>
<p>A <strong>timing ECO</strong> does not change the function at all: the logic is correct, but a few paths still violate setup or hold. The fix is local — resize a cell, insert a buffer, swap a pin, or change a cell to a different threshold-voltage variant. The nets stay the same in most cases; only cell sizes and a few added cells change.</p>
<p>A <strong>metal-only ECO</strong> is a functional ECO with a hard constraint: it can only change metal layers, because the base layers — diffusion, poly, contacts — are already manufactured into masks, or the wafers are already in the fab. The new logic has to be built entirely from spare cells placed in advance and left unconnected. This is the ECO that saves a tapeout, and it only works if someone planned for it months earlier.</p>
<p>ECOs happen for predictable reasons: a late verification bug, the last few STA violations, a spec change, or a problem found in first-silicon bring-up. Post-silicon debugging alone is about 35% of a chip's design cycle, according to one academic analysis — and the ECO is the cheapest change available at that point.</p>
<h2>Metal-only vs base-layer ECOs</h2>
<p>The dividing line in ECO work is not how big the change is. It is which masks the change touches.</p>
<p>Before any masks are made, an ECO is a <strong>pre-mask ECO</strong>: it can change anything, and the cost is engineering time, not masks. This is the cheapest ECO, and the one teams use most often during final timing closure.</p>
<p>After masks are made, the rules change completely. Every layer is formed by its own photolithographic mask, and a modern set has dozens of them — changing a layer means remaking its mask. A change that alters a transistor, adding or resizing a cell, needs new front-end masks from the bottom of the stack up: a <strong>base-layer ECO</strong>. It is expensive because the front-end masks are the most expensive ones, needing the finest writing, the most optical proximity correction and the strictest inspection. One industry tutorial puts a single critical mask at around $100,000 at 65nm, with a full set above $1 million and roughly doubling at each node.</p>
<p>A <strong>metal-only ECO</strong> avoids all of that: it changes only the wiring — the metal and via layers — and reuses transistors already on the die. The foundry holds partly processed wafers at the point where the transistors are finished and the wiring has not started, the wafer bank, and a new set of metal and via plates goes on top. Only those upper masks are remade, and the wafers skip the front-end processing entirely.</p>
<p>The cost difference is large. Published orders of magnitude put a full 28nm mask set at roughly £2–4 million against £600,000–£1.2 million for a metal-only spin; at 7nm, £8–15 million vs £2.5–5 million; at 3nm, £20–40 million vs £6–12 million. These are orders of magnitude, not quotes — real mask prices are covered by non-disclosure — but the ratio is the point: a metal-only fix costs roughly a third of a full set, and it is faster, because it skips the front-end half of the fab cycle. In the open sky130 process, for example, a metal-only fix remakes 10 of the 34 masks, while the front-end masks stay as they are.</p>
<p>The trap: resizing a cell is not a metal-only change. In the netlist it looks like a small edit, but a resized cell has different transistors, so it needs front-end masks. Only rewiring between cells that already exist is truly metal-only.</p>
<h2>Spare cells: insurance you place before tapeout</h2>
<p>A metal-only ECO can only rewire gates already on the die. That is why spare cells exist: unconnected gates — inverters, NANDs, NORs, multiplexers, flip-flops, compound AOI/OAI gates — scattered across the floorplan during physical design, so a future ECO has something to connect to. They are insurance, bought with a little silicon area before tapeout so a fix later stays in the cheap, metal-only regime.</p>
<p>Standard guidance is to budget roughly 1–5% of placed standard-cell area for spares: around 2–3% to start, toward 5% for risky or first-silicon designs, toward 1% for mature blocks. Practitioners commonly quote about 2% of total cell count, and one academic study notes CPUs can carry as many as 1% spare cells. The exact number is a judgement call, agreed with the project lead — not a script default.</p>
<p>What goes into the pool matters as much as how much of it there is. The mix should be weighted toward the workhorses — many inverters and two-input NAND and NOR gates — with a useful count of multiplexers and flip-flops, and a smaller set of AOI and OAI compound gates for denser fixes. The point is that an arbitrary future fix can be assembled from nearby spares: an AND gate is a NAND followed by an inverter. A pool of only one cell type is nearly useless.</p>
<p>How the spares are placed matters just as much: spread evenly — a uniform grid or per-region quota — so every region has spares within ECO routing reach. Spares grouped in one corner are almost as bad as no spares at all: a fix on the far side would need long wires across the die, bringing extra delay, capacitance and timing problems. A good plan is verified by a coverage check on per-region density and distribution, not just a global count.</p>
<p>Three rules keep spares usable. First, every spare input is tied to a defined level — low to ground or high to power, through a tie cell — and every spare output is left floating-safe, so unused gates draw no switching power and cannot float into an undefined state. Second, every spare is marked <code>dont_touch</code> (or <code>keep</code>), so optimisation cannot delete, resize, absorb or re-purpose it before tapeout. Third, spares are inserted after placement and before clock-tree synthesis, and downstream optimisations must preserve them — the rule most often broken by accident.</p>
<p>None of this is free — spare cells cost area and a little leakage — but against the price of a base-layer respin, it is the cheapest insurance in the flow.</p>
<h2>The ECO implementation flow</h2>
<p>A functional ECO follows a fixed sequence, and skipping a step is how ECOs go wrong. It starts with the RTL fix, verified in simulation. Then a netlist patch is created — by hand for a very small change, or with an automated ECO tool for anything larger. The tool compares the original implementation with the revised design, works out the small set of gates to add, remove or rewire, and produces a patch. Cadence Conformal ECO Designer and Synopsys Formality ECO are the two widely used tools, and both combine patch generation with formal equivalence checking.</p>
<p>The patch is then proven correct. Logic equivalence checking compares the patched netlist against the new RTL, and the result must be clean — zero non-equivalent points. This separates a real ECO from a hopeful edit: simulation cannot prove a patch correct across all input combinations, and an ECO that changes the wrong gate will pass simulation and fail in silicon.</p>
<p>Only then does the physical work begin. The patch is applied to the placed-and-routed database: new cells are placed incrementally, in free space near the logic they connect to, and only the changed nets are re-routed. Everything else — untouched placement, existing routes, the clock tree — stays exactly where it was. For a metal-only ECO, the new logic is mapped onto spare cells instead, and only metal and via layers change. Then the checks run again, all of them: freeze the design, edit only what must change, re-route incrementally, re-verify everything.</p>
<h2>Timing closure on a changed netlist</h2>
<p>Timing ECOs follow a tighter loop: run STA at all corners and modes, collect the violations, choose fixes, edit the netlist, place new cells near the old ones, route the changed nets, extract the changed parasitics, and run STA again — repeating until no violation remains, re-checking all corners every time.</p>
<p>In practice the work is generated in the signoff STA tool and handed to the place-and-route tool: signoff timing decides which cells to resize and where to add buffers, because it sees the real extracted parasitics; the physical tool makes those edits and re-routes only the touched nets. The standard commands split the work by intent: <code>fix_eco_timing</code> repairs setup and hold by resizing cells and inserting or removing buffers; <code>fix_eco_drc</code> repairs design-rule violations such as maximum transition, maximum capacitance and noise; and <code>fix_eco_power</code> recovers power once timing is met, mainly by swapping cells to a higher-threshold, lower-leakage variant where there is slack, or removing unneeded buffers.</p>
<p>Two disciplines make timing ECOs converge instead of oscillating. First, fix with a margin, not to zero: closing to a small positive slack means the change survives re-analysis after extraction, instead of passing in one run and failing in the next. Second, restrict the tool to the endpoints that actually violate — fixing paths that already pass only adds changes, and every change is a chance to break something else.</p>
<p>It is also worth knowing what a timing ECO cannot do. A worked example on a small routed multiplier shows the limit: a timing ECO that upsized nine cells, adding 0.2% more area, cut total negative slack by only 6%, because it could help endpoints close to passing but not ones with a structural problem. An ECO cannot fix a structural timing problem — a path that is too deep, two blocks placed too far apart, a clock architecture that fights the design. If the violations are structural, the honest answer is to go back further in the flow, not to keep ECO-ing.</p>
<h2>Verification and signoff after an ECO</h2>
<p>An ECO is not finished when the layout looks right. It is finished when it passes the same signoff checks as the original design, run again on the changed design.</p>
<p>For a functional ECO, the first check is equivalence: the patched netlist must be formally equivalent to the new RTL. Then timing at all corners and modes, setup and hold, including crosstalk. Then design rules on the changed nets, and the physical checks: the new cells fit, the routes pass, the power grid is intact, and antenna, IR drop and electromigration still pass. Power-intent checks run again if the design has power domains, scan-chain integrity is checked, and the changed logic goes through gate-level simulation.</p>
<p>A timing ECO needs the same list minus the strictest equivalence step — but even there, equivalence checking is worth doing, because STA does not look at function. A wrongly connected cell breaks the logic silently, and timing analysis will happily report that the broken logic meets timing: the classic ECO failure mode, and entirely avoidable.</p>
<p>ECOs are not free, even pre-mask: each round of checks costs compute and engineer time, and each round can find something new. Budget for two or three rounds, not one, and freeze the design between rounds — an ECO on top of an ECO is how teams lose track of what was verified against what.</p>
<h2>When an ECO is not enough</h2>
<p>There are four situations where the right answer is not an ECO.</p>
<p>First, when the fix needs logic that does not exist on the die. A metal-only ECO can only build from spare cells and cells freed by the change. If the bug needs twenty gates and the nearby pool has three, no amount of clever routing helps: the fix needs new transistors, and that means new base-layer masks.</p>
<p>The second is when the fix needs a cell to change size or drive strength. A resize looks like a small netlist edit but is a front-end mask change on silicon — and if the timing fix needs bigger cells where no bigger spare exists, the metal-only route is closed.</p>
<p>The third is when the problem is structural: a path too deep, a block placed too far from its partner, a clocking scheme that cannot meet skew. These are architecture and floorplan problems, and an ECO polishes them at best — the multiplier example above gained only 6% when the design needed far more. A re-spin with a floorplan or architecture change is the honest fix.</p>
<p>The fourth is when the ECOs keep coming. One ECO is surgery; five ECOs are a redesign done badly. Each ECO adds wires, consumes spares, and perturbs neighbouring timing. After several rounds the design is a patchwork, the spare pool is exhausted, and the schedule saved by avoiding a respin has been spent on ECO rounds. A clean re-spin is then cheaper, faster and safer. Knowing where that line is — and saying so in a review, when the pressure is to keep patching — is one of the more valuable judgements a PD engineer develops.</p>
<p>ECOs exist because late changes are inevitable, and the flow is built to absorb them cheaply: plan the spare cells early, keep the change minimal, prove it equivalent, re-verify everything, and know when to stop. The teams that tape out cleanly are rarely the ones that never need an ECO — they are the ones that planned for the ECO they knew was coming.</p>
<h2>Sources</h2>
<ul>
<li>ECO flow, pre-mask vs post-mask ECOs, spare cells, and timing ECO commands — SiliconPath, "The ECO Flow": <a href="https://jobs.siliconpath.in/learn/physical-design/eco-flow">https://jobs.siliconpath.in/learn/physical-design/eco-flow</a></li>
<li>ECO types, timing ECO loop, post-route ECO steps, and the worked multiplier example — ChipVerify, "ECOs After Routing": <a href="https://chipverify.com/physical-design/ecos-after-routing">https://chipverify.com/physical-design/ecos-after-routing</a></li>
<li>Timing ECOs, the STA-driven fix loop, and post-ECO checks — ChipVerify, "Timing ECOs": <a href="https://chipverify.com/static-timing-analysis/timing-ecos">https://chipverify.com/static-timing-analysis/timing-ecos</a></li>
<li>ECO definition, functional vs timing vs metal-only ECOs, and Conformal/Formality ECO tools — The Semiconductor Hub, "ECO (Engineering Change Order)": <a href="https://www.thesemiconhub.com/glossary/eco">https://www.thesemiconhub.com/glossary/eco</a></li>
<li>Metal-only vs full respin, sky130 mask counts (10 of 34 masks), and revision naming (A0/A1/B0) — ChipVerify, "First Silicon and Respins": <a href="https://chipverify.com/physical-design/first-silicon-and-respins">https://chipverify.com/physical-design/first-silicon-and-respins</a></li>
<li>Mask set cost orders of magnitude by node (28nm, 7nm, 3nm) and the wafer-bank explanation — Lizard-Spock, "Digital Design 01: Mask sets, and how a transistor gets built": <a href="http://lizard-spock.co.uk/digital-design-01-mask-sets-and-how-a-transistor-gets-built.html">http://lizard-spock.co.uk/digital-design-01-mask-sets-and-how-a-transistor-gets-built.html</a></li>
<li>Conformal ECO Designer: equivalence checking, ECO patch generation, spare-gate mapping, post-mask ECO — Cadence, "Conformal ECO Designer": <a href="https://www.cadence.com/en_US/home/resources/datasheets/conformal-eco-designer-ds.html">https://www.cadence.com/en_US/home/resources/datasheets/conformal-eco-designer-ds.html</a></li>
<li>Pre-mask Conformal ECO flow, LEC between golden and revised netlists, patch optimisation — EDN Asia, "A primer on engineering change order (ECO) using Conformal": <a href="https://www.ednasia.com/a-primer-on-engineering-change-order-eco-using-conformal/">https://www.ednasia.com/a-primer-on-engineering-change-order-eco-using-conformal/</a></li>
<li>Spare-cell density guidance (1–5% of standard-cell area), spare-cell mix, tie-offs, dont_touch preservation, spatial distribution — vibe-ic Design-for-ECO methodology: <a href="https://github.com/jueshi/vibe-ic/blob/HEAD/vibe-ic-marketplace/plugins/vibe-ic/skills/design-for-eco/SKILL.md">https://github.com/jueshi/vibe-ic/blob/HEAD/vibe-ic-marketplace/plugins/vibe-ic/skills/design-for-eco/SKILL.md</a></li>
<li>Spare-cell analysis, post-silicon debug share of design cycle (<del>35%), and spare-cell placement — University of Michigan, "Reap What You Sow: Spare Cells for Post-Silicon Metal Fix": <a href="https://web.eecs.umich.edu/">https://web.eecs.umich.edu/</a></del>valeria/research/publications/ISPD08SimSynth.pdf</li>
<li>ECO implementation flow steps, equivalence checking, and post-ECO physical checks — Digital Electronics blog, "Texas Instruments (TI) Interview Questions on Timing": <a href="https://blog.digitalelectronics.co.in/2008/12/texas-instruments-ti-interview.html?m=1">https://blog.digitalelectronics.co.in/2008/12/texas-instruments-ti-interview.html?m=1</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[The Economics of Node Migration: What Shrinking Actually Costs PD]]></title><description><![CDATA[By Pradeep Kumar Adusumalli Every few years, the same meeting happens. Someone puts up a slide showing the next process node: more density, less power…]]></description><link>https://thetapeoutlog.hashnode.dev/the-economics-of-node-migration-what-shrinking-actually-costs-pd</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/the-economics-of-node-migration-what-shrinking-actually-costs-pd</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Mon, 09 Nov 2026 05:04:24 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/35a30a43-a9f9-41ab-bd72-60c29c7b6a69.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>By Pradeep Kumar Adusumalli</em></p>
<p>Every few years, the same meeting happens. Someone puts up a slide showing the next process node: more density, less power, higher speed. The numbers look like a free lunch. Then the physical design team quietly starts doing the maths — and the lunch turns out to be the most expensive meal on the menu.</p>
<h2>Why teams migrate — and what the brochure promises</h2>
<p>The reasons are real. A new node buys density — a smaller die or more function in the same die. It buys power efficiency — battery life in a handset, thermal headroom in a server. And it buys speed — a frequency target the old node could not reach.</p>
<p>The foundry's own numbers make the case plainly. TSMC's N3E, its optimised 3nm node, offers roughly 20% more speed, over 30% power savings, and about 1.6x logic density compared with its 5nm node. The step before that was bigger: 5nm delivered about 1.8x density over 7nm. These are genuine, silicon-proven gains — and for a flagship chip they can decide a design win.</p>
<p>But notice the shape of that curve. The speed gain at the same power has narrowed to 10–15%, and each generation now buys less while costing far more. That is the whole story of node migration economics: shrinking returns on one side, exploding costs on the other.</p>
<h2>The price tag before you touch a layout</h2>
<p>The most widely quoted numbers in the industry come from International Business Strategies (IBS), and they are sobering. Designing a large SoC from scratch costs roughly $48 million at 28nm. At 7nm that becomes about $249 million. At 5nm, about $449 million. At 3nm, about $581 million. At 2nm, IBS projects about $725 million.</p>
<p>Masks are the most visible single line item. A full mask set costs on the order of $800,000 at 28nm, $3–5 million at 7nm, $5–8 million at 5nm, and $10–20 million at 3nm, according to industry estimates compiled by Silicon Analysts. A respin is not a rounding error: a partial mask rework alone runs into millions, and one modelled estimate puts the combined cost of a 3nm respin — new masks plus the schedule slip it causes — at over $100 million. IBS data also suggests first-tapeout failure rates run between 15% and 35%, so a respin is a contingency to budget for, not a remote risk to dismiss in planning.</p>
<p>Then there are the wafers. A 300mm wafer costs roughly $3,000 at 28nm, about $9,500 at 7nm, about $18,500 at 5nm, and about $19,500 at 3nm — more than six times the 28nm wafer, for about 1.6x the logic density of the previous node. The per-transistor cost, which fell reliably for decades, has stopped falling at the leading edge. Industry analysis of cost-per-transistor scaling shows it historically dropped to about 69% of the previous value with each node; at the latest nodes it is flat or rising. The shrink no longer pays for itself in silicon cost. It has to pay for itself in what the product can do — and that is a much harder bar to clear.</p>
<h2>What actually has to be redone in PD</h2>
<p>Here is the part non-PD people most often underestimate: almost everything. There is no button that takes a 5nm GDS and produces a 3nm GDS. Every layer of the physical implementation is rebuilt from the ground up.</p>
<p>The floorplan goes first. Cell heights change, macro aspect ratios change, the power grid pitch changes, and the block that fit neatly at the old node may not fit at all at the new one. Memory macros — often the largest blocks on the die — come back with different shapes and timing behaviour, so the floorplan you spent weeks converging has to be rethought around them. Analog and mixed-signal blocks are worse: they are frequently hand-redrawn, because their layout depends on device physics that changed with the process.</p>
<p>Placement follows, and it is not a re-run of the old script. The standard-cell library is new, with different drive strengths, leakage characteristics, and cell architectures. The timing models change, the parasitics change, and the constraints you carried over need re-derivation, not just re-application. A design that closed timing comfortably at the old node can arrive at the new node with fresh violations in places you did not expect — and the fixes are different, because the knobs are different.</p>
<p>The clock tree is rebuilt. Clock power, skew targets, and insertion delay all shift with the new parasitics, and the clock gating and buffering strategy that worked before may not be optimal — or even legal — under the new rules. Routing is rebuilt too, under a metal stack with different layer counts, pitches, and via rules. Via strategy itself becomes a design problem at advanced nodes: via resistance is a larger fraction of path delay, via doubling and via farms cost area, and via-related electromigration rules can quietly eat the slack on your most critical paths.</p>
<p>Then comes timing closure, which is where the schedule really goes. Signoff corners multiply: where a 28nm design might sign off across 5–7 process-voltage-temperature corners, a 3nm design can require 20–30 or more, according to SemiAnalysis' EDA market primer. Every corner is a full timing run, and every corner can produce its own violations. Variation — on-chip, across the wafer, across voltage and temperature — is a bigger fraction of the budget at every step down, so the margins you used to absorb with a little extra buffering no longer stretch far enough. The final 5% of closure, the part that always takes the longest, takes longer still.</p>
<p>None of this is a complaint about the tools. It is the physics and the economics of the node. But it is worth saying plainly to anyone who has not lived it: the RTL may survive a migration largely intact. The physical design does not. The PD team effectively designs the chip twice.</p>
<h2>IP requalification: the hidden tax</h2>
<p>If the standard cells and memories are rebuilt, the IP blocks are requalified — and that is where a surprising share of the migration budget disappears. Every licensed block in the design — processors, interfaces, memory controllers, PHYs, security subsystems — was characterised, verified, and signed off for the old node. None of that work carries over automatically.</p>
<p>Each IP block needs a new version for the new node, which means new licensing, new integration, new timing models, and new verification. Some blocks are not available on the new node at all when you need them, which forces either a wait, a redesign, or a substitute — each with its own schedule cost. Analog and mixed-signal IP is the worst of it: a PLL, an ADC, or a high-speed SerDes that was silicon-proven at the old node may need a full redesign at the new one, because its behaviour depends on transistor characteristics that do not scale neatly. There are documented cases in the industry of a single analog IP block needing many tapeouts over several years before it was validated on a new process.</p>
<p>This is also where intermediate nodes earn their keep. Foundries offer half-steps with backward-compatible design rules: TSMC's N6 is backward-compatible with N7 design rules, device models, and IP, and its N4 and N4P nodes are design-rule compatible with N5. These exist precisely because a full-node jump is so costly. If your product can get what it needs from an intermediate node, that is often the smartest money in the whole decision — most of the gain, a fraction of the requalification pain.</p>
<h2>New DRC, DFM, and rules complexity</h2>
<p>The rule deck is where the new node's difficulty becomes concrete, and it grows faster than most engineers expect. At 3nm, foundries impose more than 25,000 design rules, according to SemiAnalysis — each one a manufacturing constraint that must be satisfied simultaneously. Some counts of base-layer rules alone run higher still. For comparison, a 14nm ruleset held roughly 5,000 base-layer rules and a 7nm ruleset roughly 15,000. The rules do not just multiply; they change character. At older nodes, design rules were mostly independent geometric checks: minimum width, minimum spacing, minimum enclosure. At advanced nodes, rules are context-dependent — what is legal depends on what sits next to it, on the pattern around it, and on how the lithography will actually print it.</p>
<p>EUV lithography, multi-patterning colouring, and restrictive design rules turn layout into a constrained puzzle. Critical layers are gridded onto fixed pitches, so only a limited vocabulary of shapes is legal in the first place. DRC stops being a checklist you run at the end and becomes something you design against from the first placement. Physical verification runs get longer — a 7nm chip needs roughly 10–50x more verification compute than a 28nm chip of equivalent gate count — and 3nm tooling costs roughly 3–5x what 28nm tooling costs. None of this is optional: the foundry mandates the signoff flow, and the tapeout does not happen until every rule passes.</p>
<p>For the PD engineer, the practical consequence is a change in how the job feels. At mature nodes, a good engineer could hold most of the rules in their head and fix violations by judgement. At advanced nodes, nobody can. You work through the tool, you iterate, and you learn to distrust any fix that is not verified across all corners.</p>
<h2>The cost drivers, all in one place</h2>
<p>Pull it together and the migration bill has five big lines:</p>
<p><strong>Masks and silicon.</strong> $10–20 million for a 3nm mask set, wafers at roughly $19,500 each, and a respin exposure north of $100 million once schedule is included.</p>
<p><strong>EDA and compute.</strong> Tool costs several times higher than at mature nodes, verification compute an order of magnitude higher, and a full advanced-node design cycle of 12–24 months, mostly verification.</p>
<p><strong>Engineering effort.</strong> Bigger teams, longer schedules, more specialisation. Timing closure across 20–30 corners, DRC closure against 25,000+ rules, and IP requalification all consume senior engineers — the scarcest resource on any PD team.</p>
<p><strong>Schedule risk.</strong> A slip at an advanced node is not just a delay; it is a revenue event. One modelled estimate for a 3nm server chip put a 7-month slip at roughly $104 million in lost launch revenue and team burn. Missing a market window can cost more than the entire design budget.</p>
<p><strong>Yield ramp.</strong> Early yields at a new node are lower, and the ramp is longer. The first silicon is rarely the volume silicon, and the wafers you burn learning the process are wafers you paid leading-edge prices for.</p>
<h2>When a shrink pays — and when it just moves the pain</h2>
<p>So when is migration the right call? The honest answer is: when the product's economics demand it, and the volume supports it.</p>
<p>A shrink pays when the product genuinely needs what only the new node offers: the density to fit a bigger neural engine, the power efficiency to hold a thermal envelope, the frequency to hit a performance target that wins a socket. Flagship mobile chips and large AI accelerators live here — they ship in tens of millions of units, or sell at prices that absorb a $581 million design cost. For them, the migration bill is large, but the alternative, losing the market, is larger.</p>
<p>A shrink pays when the volume maths works. The design cost is fixed; the silicon saving is per unit. The migration has to earn back its fixed cost across the units you will actually ship, at the margin those units earn. A chip shipping a hundred thousand units a year will never earn back a leading-edge design cost, however elegant the shrink. A chip shipping fifty million units might earn it back in a quarter. This is the calculation that should happen before the floorplan starts — and the one most often skipped.</p>
<p>A shrink does not pay when it is done for the number on the slide. If the product is pad-limited rather than core-limited, density buys nothing. If the bottleneck is memory bandwidth, a faster core buys nothing. If the analogue content dominates the die, the new node buys little and costs much, because analogue barely scales. And if the schedule cannot absorb a 12–24 month design cycle plus a yield ramp, the "faster" node delivers a later product.</p>
<p>This is also why the mature nodes are not dying. The legacy-chip foundry market was worth about $45 billion in 2025 and is projected to reach about $66 billion by 2032. Around 80% of the chips in a modern vehicle are built on mature nodes of 28nm and above. Automotive, industrial, power, and connectivity chips stay on older nodes for years because the economics are better there: cheaper masks, cheaper wafers, proven yields, stable IP, and design cycles a team can actually predict. Staying put is not falling behind. For most products, it is the rational choice.</p>
<p>And there is a third path reshaping the whole decision: chiplets. Instead of moving the entire design to the new node, you move only the part that benefits — the compute tile — and leave the I/O, analogue, and memory interfaces on a cheaper, mature node, joining them in one package. When the cost per transistor stops falling at the leading edge, there is no advantage in paying new-node design costs for blocks that do not benefit from the new node's performance. Chiplets let each block live on the node that suits it — and change the question from "do we move?" to "what do we move?"</p>
<h2>The verdict</h2>
<p>Node migration is not a technology decision. It is a business decision with a technology bill attached, and the PD team pays most of that bill in effort and schedule. The brochure numbers are real, but they are the gross return, not the net one. The net return is what is left after the mask set, the wafers, the tooling, the requalification, the 25,000 rules, the 20–30 signoff corners, the longer cycle, and the respin risk.</p>
<p>For a PD engineer asked whether to migrate, the useful questions are these. What exactly does the product gain that it cannot get any other way? How many units will pay for the fixed cost? Is the design core-limited or pad-limited, compute-bound or memory-bound? Can an intermediate node, or a chiplet split, buy most of the gain for a fraction of the pain? And is the schedule honest about closure, requalification, and yield ramp — or is it a date someone needed to be true?</p>
<p>Answer those honestly and the decision usually makes itself. Sometimes the answer is to migrate, eyes open, with the budget and schedule to match. Sometimes it is to stay, and spend the money on architecture, packaging, or software instead. Either way, the worst outcome is migrating because the slide said to, and discovering the real price only when the tapeout is already late.</p>
<p>The shrink is not free. It never was. It is just that, for a long time, it was cheap enough that nobody had to ask what it cost. That time is over — and the PD team is the one holding the receipt.</p>
<h2>Sources</h2>
<ul>
<li>International Business Strategies design-cost figures (28nm ~$48M, 7nm ~$249M, 5nm ~$449M, 3nm ~$581M, 2nm ~$725M projected), via SemiWiki: <a href="https://semiwiki.com/artificial-intelligence/375867-the-725-million-chip-development-why-execution-risk-not-ip-cost-now-dictates-advanced-node-survival/">https://semiwiki.com/artificial-intelligence/375867-the-725-million-chip-development-why-execution-risk-not-ip-cost-now-dictates-advanced-node-survival/</a></li>
<li>IBS design-cost figures and the "silicon ceiling" argument (28nm $40M, 7nm $217M, 5nm $416M, 3nm $590M), via Chips &amp; Change: <a href="https://chipsandchange.tech/silicon-ceiling-why-change-no-longer-optional/">https://chipsandchange.tech/silicon-ceiling-why-change-no-longer-optional/</a></li>
<li>TSMC advanced technology node characteristics — N3E vs N5 (<del>20% speed, &gt;30% power savings, ~1.6x logic density), N5 vs N7 (</del>1.8x density), N4/N6 design-rule compatibility: <a href="https://www.tsmc.com/english/dedicatedFoundry/technology/platform_smartphone_tech_advancedTech">https://www.tsmc.com/english/dedicatedFoundry/technology/platform_smartphone_tech_advancedTech</a></li>
<li>SemiAnalysis EDA Market Primer — 25,000+ design rules at 3nm, PVT corners 5–7 at 28nm to 20–30+ at 3nm, 3nm EDA tool costs 3–5x 28nm, 12–24 month design cycles, 10–50x verification compute at 7nm vs 28nm: <a href="https://newsletter.semianalysis.com/p/eda-market-primer">https://newsletter.semianalysis.com/p/eda-market-primer</a></li>
<li>Silicon Analysts semiconductor manufacturing cost guide — wafer costs by node (28nm ~$3,000, 7nm ~$9,500, 5nm ~$18,500, 3nm ~$19,500) and design costs by node: <a href="https://siliconanalysts.com/guide/semiconductor-costs">https://siliconanalysts.com/guide/semiconductor-costs</a></li>
<li>Silicon Analysts foundry engagement guide — full mask set costs by node (28nm ~$800K, 7nm $3–5M, 5nm $5–8M, 3nm $10–20M): <a href="https://siliconanalysts.com/guide/foundry-engagement">https://siliconanalysts.com/guide/foundry-engagement</a></li>
<li>First-tapeout failure rates (15–35%) and tapeout costs ($40–50M at 5nm), via IBS data cited in this analysis of the semiconductor chain: <a href="https://dan.rio/en/blog/semiconductors-and-sovereignty.html">https://dan.rio/en/blog/semiconductors-and-sovereignty.html</a></li>
<li>Cost-per-transistor scaling — historical ~69% per node, now flat or rising at leading-edge nodes; chiplet node-mixing rationale, via Chiplet Marketplace: <a href="https://chiplet-marketplace.com/insights/expert-perspectives/introduction-to-chiplets-why-the-industry-is-moving-beyond-monolithic-designs">https://chiplet-marketplace.com/insights/expert-perspectives/introduction-to-chiplets-why-the-industry-is-moving-beyond-monolithic-designs</a></li>
<li>Multi-die design complexity and rising cost per transistor at successive leading-edge nodes, via SemiEngineering: <a href="https://semiengineering.com/multi-die-design-pushes-complexity-to-the-max/">https://semiengineering.com/multi-die-design-pushes-complexity-to-the-max/</a></li>
<li>Legacy/mature-node foundry market — $45 billion in 2025, projected $66 billion by 2032: <a href="https://pmarketresearch.com/it/legacy-chips-wafer-foundry-market/">https://pmarketresearch.com/it/legacy-chips-wafer-foundry-market/</a></li>
<li>Mature-node share of automotive chips (~80% of chips in vehicles on nodes of 28nm and above) and mature-node capacity constraints: <a href="https://accuristech.com/blog/chips-act-semiconductor-reshoring-component-availability/">https://accuristech.com/blog/chips-act-semiconductor-reshoring-component-availability/</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[Timing Budgets: How the Top Level Hands Slack to Every Block]]></title><description><![CDATA[Pradeep Kumar Adusumalli Every block in a large chip closes timing on its own, on its own schedule, with its own team. The top level never sees the inside…]]></description><link>https://thetapeoutlog.hashnode.dev/timing-budgets-how-the-top-level-hands-slack-to-every-block</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/timing-budgets-how-the-top-level-hands-slack-to-every-block</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Sun, 08 Nov 2026 05:18:30 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/c6972f0a-a7c1-488e-b838-d7dca9e268a5.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Pradeep Kumar Adusumalli</em></p>
<p>Every block in a large chip closes timing on its own, on its own schedule, with its own team. The top level never sees the inside of those blocks — it sees only their ports. So before any block is built, someone has to decide, in advance, how much of the clock period each block is allowed to spend. That decision is the timing budget, and it is the single most consequential guess in a hierarchical flow. Get it right and blocks close independently and stitch together cleanly. Get it wrong and every block can report green while the chip fails at the top.</p>
<h2>Why the Top Level Has to Guess First</h2>
<p>A large design cannot be analyzed as one flat netlist in a single run: the memory and runtime are too large, and several teams work on different parts at the same time. So the chip is divided into blocks, each block is analyzed on its own, and in the top-level run each finished block is replaced by a smaller timing model. That is the whole idea of hierarchical STA.</p>
<p>The price of this division is that the blocks have to agree on the timing at their boundaries before they exist in their final form. A block team cannot wait until the neighbouring block is routed to know when data will arrive at its input pins. The budget is that agreement, written down before the blocks are built, so that each team can work independently.</p>
<h2>What a Budget Actually Is</h2>
<p>A budget splits the clock period of a path that crosses block boundaries. Take a path that starts in block A, passes through the top level, and ends in block B, with a 2.0 ns clock period. Suppose block A is given 0.9 ns, the top-level wires 0.2 ns, and block B 0.7 ns. The three shares add up to 1.8 ns, leaving 0.2 ns as margin.</p>
<p>Each block team receives its share as constraints at its ports: an output delay for the part of the path that lies after the block, and an input delay for the part that comes before it. In SDC these are <code>set_output_delay</code> and <code>set_input_delay</code>, and they are chosen so that the blocks can be designed independently. Block A, for example, gets an output delay of 0.9 ns — the 0.2 ns of top-level wire plus the 0.7 ns inside block B — because that is everything that must still happen after its data leaves the port.</p>
<p>A budget, then, is not a measurement. It is a first guess about delays that do not exist yet. As blocks get implemented and the real delays become known, the budgets are updated, and the loop between the top level and the blocks repeats.</p>
<h2>How Budgets Are Set in Practice</h2>
<p>Good budgeting starts with the floorplan, not with a spreadsheet. You need the block shapes, the pin locations, and the top-level timing constraints before you can divide anything sensibly. If a globally routed design is available, better still — the single hardest quantity to predict before layout is the delay of the global interconnect between blocks.</p>
<p>The budgeting step itself is usually done by a tool that distributes slack rather than inventing delays. It analyzes the inter-block timing arcs, then writes out the block-level SDC: input and output delays, clock constraints, and the timing exceptions that apply at the boundary. One important detail: the budgeter does not change the timing of the design. It only divides the available slack among the blocks. That is why the quality of the inputs matters so much — good budgeting inputs produce SDC that is neither over- nor under-constrained, and bad inputs produce a budget no block team can trust.</p>
<p>Two classic mistakes frame the whole exercise. Under-budgeting gives a block less time than its own cells need: if a block's internal cell delay alone already exceeds its share, no amount of careful routing inside the block will fix it. Over-budgeting does the opposite — it hands a block more time than it needs and wastes slack that another block is short of. In fact the two are often the same mistake seen from two sides: one block's over-budget is frequently the reason another block is under-budgeted, because the total available time is fixed by the specification.</p>
<h2>Where the Margin Goes</h2>
<p>The 0.2 ns left over in the example above is not spare time waiting to be spent. It is the reserve that absorbs everything the budget could not predict. In practice it covers several distinct risks, and it helps to name them separately instead of lumping them into one number:</p>
<ul>
<li><strong>Clock uncertainty.</strong> Before clock-tree synthesis, the skew between two flops is unknown, so the budget carries an uncertainty margin on the clock. After CTS, uncertainty still covers jitter — the cycle-to-cycle variation of the clock edge itself.</li>
<li><strong>On-chip variation.</strong> Two identical cells on the same die do not switch at exactly the same speed, so sign-off STA derates delays: for a setup check the launch clock and the data path are made slower and the capture clock faster, and the reverse for hold. A flat derate of a few percent sounds small, but it is applied to every delay on the path, so long paths pay far more than short ones. In one routed demo design, applying a 5% flat derate moved the worst setup slack from +0.134 ns to −0.353 ns — a passing path turned into a failing one purely because variation was assumed.</li>
<li><strong>Top-level wire error.</strong> The budget guessed the inter-block wire delay before routing existed. If the route turns out longer, that error comes straight out of the margin.</li>
<li><strong>Crosstalk and signal-integrity effects</strong> on long top-level nets, which the block-level runs never see.</li>
<li><strong>ECO reserve.</strong> Some margin is deliberately held back so that late fixes — a buffer here, a reroute there — can be absorbed without reopening every block's budget.</li>
</ul>
<p>A useful rule of thumb from the budgeting flows used in industry is to keep the slack being distributed reasonable relative to the clock period — if the numbers you are handing out are already a large fraction of the period, timing closure may not be achievable at all, and the problem is the budget, not the blocks.</p>
<h2>Why a Block That Meets Its Budget Can Still Break the Chip</h2>
<p>This is the question that catches engineers new to hierarchical flows. Every block reports positive slack, every block-level run is clean — and the top level still fails. Here are the ways it happens, roughly in order of how often they bite.</p>
<p><strong>1. The interface paths are the hardest paths in the design.</strong> The timing of the paths that cross a block boundary is widely regarded as the biggest challenge in hierarchical timing closure. There are only four ways a clock and a data signal can interact across a partition — clock in with data in, clock in with data out, clock out with data out, and clock out with data in — and two of them are genuinely difficult. When the clock enters the block but the data leaves it, or the clock leaves while the data enters, a round-trip delay appears in the timing equations: the delay goes out and comes back, and both the pad delays and the internal delays have to fit in one period. The other two cases are far more forgiving, because clock and data travel in the same direction and their pad delays largely cancel. If your budget treated all four cases the same, the round-trip ones will be the ones that fail.</p>
<p><strong>2. Required and available are not the same number.</strong> At a block boundary there are two different setup times. The <em>required</em> setup time is what the block's internal flops demand, derived from the inside out. The <em>available</em> setup time is what the top level actually offers, derived from the chip-level constraint minus the clock-data skew at the top level. A block can meet its required number while the available number at the top is smaller, because the real clock-data skew at the top level differs from what the budget assumed. The same applies to hold, with the signs reversed. When the budget's skew assumption and the real skew disagree, the interface fails even though each side, checked alone, looked fine.</p>
<p><strong>3. The top level applies variation the block never saw.</strong> A block is usually closed with the derates and clock uncertainty its team was told to use. At the top level, the same path is re-checked with the full sign-off derates, the real clock tree, and clock reconvergence pessimism removal. The derate example above shows how a 5% assumption alone can flip a path. Worse, if the block was closed with one set of derates and the top level uses another, the two runs are not even measuring the same thing.</p>
<p><strong>4. The top-level wires are longer than the budget guessed.</strong> Blocks placed at the top level act as routing obstructions, and the nets between them have to be routed through the channels between blocks. Those routes tend to be long, and long routes cause both timing and signal-integrity problems. Congestion in one area can force detours that add delay no budget predicted. And unlike logic inside a block, the timing of a block's I/O cannot be changed during top-level optimization — the block is already finished, so the top level has nowhere to recover the lost time except by re-budgeting.</p>
<p><strong>5. The model is not the block.</strong> At the top level, blocks appear as timing models, not netlists. An extracted timing model (ETM) is a Liberty-format cell with timing arcs between the ports; it is fast and small, but it hides everything inside. An interface logic model (ILM) keeps the logic connected to the ports and discards the rest, so it is more accurate and bigger. Neither is exact: in one documented comparison, the setup slacks from a model matched the flat netlist to four decimal places, while a hold slack differed by 0.033 ns. That is why a full flat run of the whole design is still done as a last check before sign-off — models are approximations, and hold is where the approximation shows. A related trap is using a model built at one corner in an analysis at another corner; a timing model holds numbers for one corner only.</p>
<p><strong>6. Exceptions and modes do not propagate by themselves.</strong> False paths, multicycle paths, and mode-specific constraints defined at the top level have to reach the block SDC intact. If an exception is missing at the block level, the block team over-constrains a path that was meant to be ignored — or worse, a path the top level assumed was an exception gets timed as a real path and fails. The same applies to clocks: a budget written for one mode does not automatically hold in another.</p>
<h2>The Loop That Makes Budgets Work</h2>
<p>Because a budget is a guess, the flow around it is iterative by design:</p>
<ol>
<li>The chip planner writes budgets for every block, as input and output delays and clock latencies.</li>
<li>Each block is implemented and analyzed on its own against those constraints.</li>
<li>When a block is stable, a timing model is written for it — an ETM or an ILM — for every corner and mode.</li>
<li>The top level is analyzed with the models, which is fast and small, plus its own cells and wires.</li>
<li>Violations at the boundaries are fixed by changing the budgets or the blocks, and the loop repeats.</li>
<li>At sign-off, a full flat run of the whole design is done as a final check, because the models are approximations.</li>
</ol>
<p>The loop is the point. A budget that is never revised is a budget that is wrong; the first version exists so that implementation can start in parallel, not so that it can be frozen.</p>
<h2>A Practical Checklist for the Block Owner</h2>
<p>If you receive a budget, do not just type it into your SDC and start placing cells. Check these first:</p>
<ul>
<li><strong>Is the floorplan behind it real?</strong> Pin locations and block placement decide the wire delay your budget assumes. A budget written before the floorplan settled is a guess on top of a guess.</li>
<li><strong>Do the round-trip interfaces get the tightest numbers?</strong> Clock-in-data-out and clock-out-data-in paths deserve the most scrutiny and the most margin.</li>
<li><strong>Is the margin visible, or already spent?</strong> If the budget consumed the whole period with no reserve for derates and wire error, the first surprise will be a violation with no room to absorb it.</li>
<li><strong>Do the derates and uncertainty match the top level?</strong> Your block run and the top-level run must apply the same variation assumptions, or the two results cannot be compared.</li>
<li><strong>Are the exceptions in your SDC?</strong> Confirm that every false path and multicycle path at your boundary is present in the budget you received.</li>
<li><strong>Is the model fresh?</strong> When a neighbouring block changes, its model changes. A top-level run against a stale model is a clean report about a design that no longer exists.</li>
<li><strong>Check hold at the top, not just in the block.</strong> Model approximations show up in hold first, and hold is checked at the fast corner where everything is at its quickest.</li>
</ul>
<p>Timing budgets are, in the end, a communication device. The top level cannot build the chip and the blocks cannot see each other, so the budget is the contract that lets both sides work at the same time. Treat it as a living document — set it from a real floorplan, keep the margin explicit, name what the margin is for, and re-budget as soon as real delays arrive. The chips that close smoothly are rarely the ones with the cleverest blocks. They are the ones whose budgets were honest about what nobody knew yet.</p>
<h2>Sources</h2>
<ul>
<li>ChipVerify — Hierarchical STA: Budgets, ETMs and ILMs: <a href="https://chipverify.com/static-timing-analysis/hierarchical-sta">https://chipverify.com/static-timing-analysis/hierarchical-sta</a></li>
<li>EDN — Hierarchical timing concepts: <a href="https://www.edn.com/design/integrated-circuit-design/4423327/hierarchical-timing-concepts">https://www.edn.com/design/integrated-circuit-design/4423327/hierarchical-timing-concepts</a></li>
<li>VLSI Concepts — Hierarchical Design Flow, Part 2 (timing budgeting, under- and over-budgeting): <a href="http://www.vlsi-expert.com/2012/04/hierarchical-design-flow-part-2.html">http://www.vlsi-expert.com/2012/04/hierarchical-design-flow-part-2.html</a></li>
<li>ChipVerify — On-Chip Variation, Derates and Sign-Off Corners: <a href="https://chipverify.com/physical-design/on-chip-variation-in-physical-design">https://chipverify.com/physical-design/on-chip-variation-in-physical-design</a></li>
<li>Gursimran S. — OCV Derating &amp; CPPR Explained: <a href="https://www.guriai.com/blog/ocv-derating-cppr-physical-design.html">https://www.guriai.com/blog/ocv-derating-cppr-physical-design.html</a></li>
<li>Renesas Mobile — Timing Budgeting Flow (slack distribution, allocate_fp_budgets): <a href="https://www.scribd.com/presentation/98386181/Timing-Budgeting-Flow">https://www.scribd.com/presentation/98386181/Timing-Budgeting-Flow</a></li>
<li>VLSIWeb — STA Interview Questions: Setup, Hold, SDC and Sign-off Explained: <a href="https://vlsiweb.com/sta-interview-questions/">https://vlsiweb.com/sta-interview-questions/</a></li>
<li>BlinkNBuild — PVT Corners, OCV Derating, AOCV, POCV and CRPR: <a href="https://www.blinknbuild.in/courses/static-timing-analysis/vol11-corners-and-ocv">https://www.blinknbuild.in/courses/static-timing-analysis/vol11-corners-and-ocv</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[Inside a Standard Cell: What PD Engineers Should Know About Cell Layout]]></title><description><![CDATA[Pradeep Kumar Adusumalli Every standard cell in a design is a tiny pre-designed layout: transistors, wires, and pins packed into a fixed-height rectangle…]]></description><link>https://thetapeoutlog.hashnode.dev/inside-a-standard-cell-what-pd-engineers-should-know-about-cell-layout</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/inside-a-standard-cell-what-pd-engineers-should-know-about-cell-layout</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Sat, 07 Nov 2026 04:04:58 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/f8f206ea-8732-4302-805e-2b26aca3f0e8.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Pradeep Kumar Adusumalli</em></p>
<p>Every standard cell in a design is a tiny pre-designed layout: transistors, wires, and pins packed into a fixed-height rectangle. To a place-and-route tool, a cell is mostly a box with pins. To the physical design engineer, that box is not a black box at all. What is inside it — how tall it is, where its pins sit, how its transistors are built — shapes placement density, routing congestion, and timing in ways that no amount of router effort can fully undo. A block that will not route cleanly is often not a routing problem. It is a cell problem that only became visible at routing.</p>
<p>This post looks inside the standard cell at the layout level: the anatomy every cell shares, what track height really means, why pin access fails, how FinFET and nanosheet devices changed cell architecture, and what to check before blaming the router.</p>
<h2>The anatomy of a standard cell</h2>
<p>A standard cell has a fixed height and a variable width. The height is the same for every cell in a library, so cells can be lined up in rows like bricks, abutting each other and sharing power rails. The width varies with the logic function and the drive strength: a two-input NAND is narrow, a flip-flop is wide, and a high-drive buffer is wider still.</p>
<p>Across the top and bottom of every cell run the power rails — VDD along one edge, VSS along the other, usually drawn in metal 1. When cells are placed in a row, the rails meet end to end and form continuous horizontal power lines, which is why the rails must sit in exactly the same position in every cell. The cell boundary, or place-and-route boundary, is drawn around the cell and defines its legal area: neighbouring cells abut at this boundary, and design rules are written so that two cells placed side by side are still clean at the shared edge.</p>
<p>Inside the boundary, the layout follows a familiar pattern. PMOS transistors sit in the upper half of the cell, NMOS in the lower half, with gates running vertically across the diffusion. Internal connections are routed mostly in metal 1, and each input and output of the cell is brought out to a pin — a small metal shape the router can connect to. The pins are the cell's only interface to the rest of the design: everything the router knows about a cell is its boundary, its pins, and the routing blockages inside it.</p>
<p>That last point matters more than it sounds. The router cannot see the transistors, and it cannot move a pin. If a pin is hard to reach, the router will spend vias and detours trying to reach it, and those detours cost routing tracks, add resistance, and create congestion that spreads to neighbouring cells. Cell layout is therefore not only the library team's problem. Its consequences land on the PD engineer's desk.</p>
<h2>Track height: what 7.5T, 9T, and 12T actually mean</h2>
<p>A cell's height is measured in tracks. One track is roughly the pitch of one routing wire — the minimum metal width plus the minimum spacing — so a "9-track" cell is nine routing tracks tall. Cell height is the number of tracks multiplied by the metal pitch, which is why the same "9T" label means different absolute heights at different nodes.</p>
<p>Track height is the first and largest tradeoff in a cell library, and it trades three things against each other:</p>
<ul>
<li><strong>Density.</strong> Shorter cells pack more logic into the same silicon area. A 7.5-track library is denser than a 9-track library built in the same process.</li>
<li><strong>Routability.</strong> Taller cells leave more metal-1 tracks free inside the cell, so there is more room for the cell's own internal routing — and, just as importantly, more room for the router to reach pins without fighting the cell's internal wires.</li>
<li><strong>Speed.</strong> Taller cells fit more and wider transistors, so they can be faster and stronger. Shorter cells squeeze the transistors and give up speed.</li>
</ul>
<p>The numbers make the tradeoff concrete. At the 7 nm generation, published work on standard-cell architectures describes two libraries side by side: a 9-track library and a 7.5-track library. The 9-track cells carry more fins per device — four fins per transistor against three in the 7.5-track cells — and provide smaller delay at the cost of higher power and area, while the 7.5-track cells are on average smaller. The same pattern shows up at older nodes: 12-track libraries are taller, with more metal-1 routing space inside the cell and faster cells, while 9-track libraries are more compact but slower. At 22/28 nm, one vendor's 6.5-track ultra-high-density library reported 8–13% area reduction and 13–26% power reduction against its mainstream libraries — bought with less room inside the cell for routing and transistors.</p>
<p>There is no universally right track height. High-density libraries suit area-constrained blocks; tall libraries suit speed-critical paths and congested blocks. The mistake is treating the library as a given and then being surprised when a dense block will not route. Track height is a design decision, and it is cheapest to make it before placement, not after the congestion map turns red.</p>
<h2>Pin access: where routing actually fails</h2>
<p>Pin access is the question of whether the router can physically reach a cell's pins. It sounds trivial — the pin is right there, on the cell — but it has become one of the hardest problems in backend design at advanced nodes. A published study from the University of Texas at Austin puts it plainly: standard-cell pin access is one of the most challenging issues in sub-14 nm physical design, because pin density has risen, the number of routing tracks has not kept up, and multiple-patterning lithography adds colouring and geometry constraints that make neighbouring pins interfere with each other.</p>
<p>The failure mode is specific. Each pin has only a limited number of access points — places where a via can land on it from the metal above — and those access points interfere with each other when pins sit close together. During detailed routing, the routes for the first few nets can block the metal-1 pin of a remaining net, so that net has no legal way in. The router has not failed to find a path; there is no path. The cell's pins were laid out in a way that makes one of them unreachable once its neighbours are connected.</p>
<p>This is why pin access is checked at the library level, before any design uses the cells. One standard method is to instantiate every cell type, place them all in a row, and try to route to every pin from outside; if routing fails, the cell has an accessibility problem. A more recent approach, presented at DAC 2025, places cells in topologies that mimic the congestion of a real SoC and runs the same router engine and rules used in actual place and route, reporting DRC violations and via counts for each cell. The point of both methods is the same: find the bad pin in the library, not in your block at 2 a.m.</p>
<p>For the PD engineer, the practical consequences of pin access show up as congestion that does not respond to the usual fixes. If a hotspot is full of cells with many pins — multi-bit flip-flops are a classic case — the problem may be that each cell demands more routing resource than its footprint suggests. The established fixes are placement-level, not routing-level: add white space between neighbouring cells so the lower metals have room for pin connections, or apply cell padding — a keep-out margin around high-pin-count cells that stops other cells being placed too close to them. Spreading cells apart feels wasteful when utilisation is the metric being watched, but a utilisation number that ignores pin access is measuring the wrong thing.</p>
<h2>Cell architecture in the FinFET and nanosheet era</h2>
<p>The transistors inside the cell have changed twice in the last decade, and each change rewrote what a cell can be.</p>
<p><strong>FinFET: drive strength in discrete steps.</strong> In a FinFET, the channel is a vertical fin, and the transistor's drive current is set by the number of fins. A cell designer cannot widen a device by a little; they can only add or subtract whole fins, so drive strength comes in discrete steps. The total device width is the number of fins multiplied by the number of fingers, and the choice is constrained by the cell height and the poly pitch. This quantisation is one reason libraries offer so many drive-strength variants of the same logic function: the granularity has to come from the library, because it cannot come from continuous sizing inside the cell.</p>
<p><strong>Nanosheet: width becomes a dial again.</strong> In a gate-all-around nanosheet device, the channel is a stack of horizontal sheets, and the sheet width can be adjusted continuously. As CEA-Leti put it when demonstrating stacked nanosheet devices, the wide range of variable nanosheet widths allows more design flexibility, which is not possible for FinFET because of its discrete number of fins. For cell design, that means drive strength can be tuned more finely inside the same cell footprint — one of the reasons the move to nanosheets at 3 nm and 2 nm is as much a cell-architecture story as a device story.</p>
<p><strong>Threshold voltage variants.</strong> Every library ships in multiple Vt flavours — typically regular, low, and high threshold voltage, with ultra-low and extra-low variants in some libraries. The rule is simple and worth stating plainly: put low-Vt cells on the critical paths and high-Vt cells everywhere leakage costs more than speed. Low-Vt is the fastest and leakiest; high-Vt is the slowest and tightest on leakage. Mixing them is how a design buys timing where it needs it and pays for it in leakage only where it must.</p>
<p><strong>Multi-height and multi-bit cells.</strong> Not every cell fits in one row. Complex cells — large flip-flops, some clock gates, wide multiplexers — are built as multi-height cells that span two or more adjacent rows, with taller cells carrying more or larger diffusion regions. Multi-bit flip-flops take the same idea further: several flops share clock and control logic in one cell, which cuts the pin count and fits more function into the same rows. Both are area wins, but both change the placement problem: a multi-height cell constrains the rows around it, and a multi-bit flop concentrates many pins in one place, which brings the pin-access problem back.</p>
<p><strong>Cell height is fixed within a library.</strong> A useful illustration comes from the open-source SkyWater 130 nm libraries: the high-density variant uses 2.72 µm tall cells, the medium-speed variant 4.8 µm, and the high-speed variant about 6.0 µm. The height is fixed within each library — you cannot mix 2.72 µm and 4.8 µm cells in the same row. Library choice is therefore a design-level decision, not something you change cell by cell, and it is why the common practice is to mix libraries: high-speed cells on critical paths, high-density cells everywhere else.</p>
<h2>What this means for physical design</h2>
<p>Most cell-related pain in a block can be traced to one of a small number of causes, and each has a check that costs less than a routing iteration:</p>
<p><strong>1. Congestion that tracks pin density, not cell density.</strong> A congestion map shows where routing demand exceeds routing supply. High cell density is one cause, but high pin density is another, and they need different fixes. If a hotspot is full of multi-bit flops or wide multiplexers, spreading cells or padding the high-pin cells is the fix — not more routing effort. One published account of congestion work notes that global-routing overflow above roughly 1% makes routing hard to complete cleanly, with shorts and DRC violations following; whether or not that exact threshold holds in your flow, the direction is right: check overflow early, before detailed routing turns it into violations.</p>
<p><strong>2. Pin-access failures hiding as DRC violations.</strong> When detailed routing finishes with a cluster of violations around a particular cell type — minimum-area violations on pins, via-enclosure problems, or shorts at the cell edge — suspect the cell before the router. Reproduce it the way library teams do: place the cell in a test row and try to route to every pin. If it fails there, no placement will save it, and the fix belongs to the library, not to your floorplan.</p>
<p><strong>3. Utilisation targets that ignore the library.</strong> A 95% utilisation target means something very different in a 7.5-track library than in a 12-track one, and different again when the block is full of multi-bit flops. Utilisation is a planning number, not a physical law. If the block is congested at a utilisation the plan says should be fine, the plan's assumption about the library was wrong, and the honest fix is to lower the target, change the library mix, or spread the cells — not to keep routing until the violations disappear.</p>
<p><strong>4. Drive-strength granularity on critical paths.</strong> In a FinFET library, the next drive strength up is a whole fin, and the timing gain comes with a real area and power cost. If a critical path needs a strength that does not exist in the library, no amount of placement tuning will invent it. Check the library's drive-strength steps before assuming the path can be closed, and remember that nanosheet libraries — where they are available — offer finer steps for exactly this reason.</p>
<p><strong>5. Vt mix as a placement problem.</strong> Multi-Vt optimisation swaps cells after placement, and a late swap can change the congestion picture: low-Vt cells are often larger, and a block that routed cleanly with high-Vt cells can become congested after a Vt swap on critical paths. Re-check congestion after Vt optimisation, and treat the Vt mix as part of the placement problem rather than a post-placement polish.</p>
<h2>The takeaway</h2>
<p>A standard cell is a compromise frozen in layout: a fixed height chosen against density, routability, and speed; pins placed where the library designer thought the router could reach them; transistors sized in whole fins or, in the nanosheet era, in continuously variable sheet widths; threshold voltages and drive strengths offered as a menu the design must choose from. None of those choices is visible in the netlist, and all of them shape what placement and routing can achieve. The PD engineer who understands what is inside the cell — track height, pin access, and cell architecture — can read a congestion map as a story about the library rather than a mystery about the router, and can fix the cause instead of the symptom. The cell is the smallest unit of the design. It is also where most of the design's routability is decided.</p>
<h2>Sources</h2>
<ul>
<li>Standard cell guidelines for analog layout (cell height, PR boundary, pin placement): <a href="https://www.slideshare.net/slideshow/standard-cell-guidelines-for-analog-layout/284995066">https://www.slideshare.net/slideshow/standard-cell-guidelines-for-analog-layout/284995066</a></li>
<li>VLSI Junction — 12 Track vs 9 Track Standard Cell: <a href="http://www.vlsijunction.com/2015/10/12-track-vs-9-track-standard-cell.html">http://www.vlsijunction.com/2015/10/12-track-vs-9-track-standard-cell.html</a></li>
<li>EE Times Asia — 6.5-Track Ultra High Density Standard Cells on 22/28 nm: <a href="https://www.eetasia.com/downloads/6-5-track-ultra-high-density-standard-cells-on-22-28nm/">https://www.eetasia.com/downloads/6-5-track-ultra-high-density-standard-cells-on-22-28nm/</a></li>
<li>Sherazi et al. — Architectural strategies in standard-cell design for the 7 nm and beyond technology node (SPIE Journal of Micro/Nanolithography, MEMS, and MOEMS, 2016): <a href="https://www.spiedigitallibrary.org/journals/journal-of-micro-nanolithography-mems-and-moems/volume-15/issue-1/013507/Architectural-strategies-in-standard-cell-design-for-the-7nm-and/10.1117/1.JMM.15.1.013507.full">https://www.spiedigitallibrary.org/journals/journal-of-micro-nanolithography-mems-and-moems/volume-15/issue-1/013507/Architectural-strategies-in-standard-cell-design-for-the-7nm-and/10.1117/1.JMM.15.1.013507.full</a></li>
<li>Arm — Standard Cell Design and Optimization Methodology for ASAP7 PDK (ICCAD 2017 tutorial): <a href="https://community.arm.com/cfs-file/__key/communityserver-blogs-components-weblogfiles/00-00-00-37-98/0160.iccad2017_5F00_asap7_5F00_library.pdf">https://community.arm.com/cfs-file/__key/communityserver-blogs-components-weblogfiles/00-00-00-37-98/0160.iccad2017_5F00_asap7_5F00_library.pdf</a></li>
<li>University of Texas at Austin (UTDA) — Standard Cell Pin Access and Physical Design in Advanced Lithography: <a href="https://www.cerc.utexas.edu/utda/publications/C187.pdf">https://www.cerc.utexas.edu/utda/publications/C187.pdf</a></li>
<li>DAC 2025 — Automated Topology based Pin Access Checker for Correct by Construction Standard Cells Design: <a href="https://62dac.conference-program.com/presentation/?id=ENGPOST097&amp;sess=sess264">https://62dac.conference-program.com/presentation/?id=ENGPOST097&amp;sess=sess264</a></li>
<li>Cadence Community — How to analyze pin accessibility?: <a href="https://community.cadence.com/cadence_technology_forums/f/digital-implementation/1229/how-to-analyze-pin-accessibility">https://community.cadence.com/cadence_technology_forums/f/digital-implementation/1229/how-to-analyze-pin-accessibility</a></li>
<li>Design-Reuse — Enhancing VLSI Design Efficiency: Tackling Congestion and Shorts with Practical Approaches and PnR Tool (ICC2): <a href="https://us.design-reuse.com/articles/?id=54881&amp;print=yes">https://us.design-reuse.com/articles/?id=54881&amp;print=yes</a></li>
<li>Silvaco — Standard Cell Library IP: Multi-VT and Low Power: <a href="https://silvaco.com/products/design/design-ip/foundation-ip/standard-cell/">https://silvaco.com/products/design/design-ip/foundation-ip/standard-cell/</a></li>
<li>ChipVerify — Library Variants (RTL Synthesis): <a href="https://chipverify.com/rtl-synthesis/library-variants">https://chipverify.com/rtl-synthesis/library-variants</a></li>
<li>US Patent 11,776,950 — Integrated circuit including cells with different heights: <a href="https://patents.justia.com/patent/11776950">https://patents.justia.com/patent/11776950</a></li>
<li>US Patent US11271011B2 — High performance standard cell design techniques in FinFET based library using local layout effects: <a href="https://patents.google.com/patent/US11271011B2/en">https://patents.google.com/patent/US11271011B2/en</a></li>
<li>CEA-Leti / Electronic Specifier — Architecture for HPC devices using nanosheet fabrication: <a href="https://www.electronicspecifier.com/design-automation/architecture-for-hpc-devices-using-nanosheet-fabrication">https://www.electronicspecifier.com/design-automation/architecture-for-hpc-devices-using-nanosheet-fabrication</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[Memory Is Now Redesigning the Chip: HBM Fabs, Long-Term Deals, and a New Memory Architecture — This Week in Semiconductors]]></title><description><![CDATA[Pradeep Kumar Adusumalli Last week the industry's story was that memory is sold out for years. This week the industry answered — and the answer is…]]></description><link>https://thetapeoutlog.hashnode.dev/memory-is-now-redesigning-the-chip-hbm-fabs-long-term-deals-and-a-new-memory-architecture-this-week-in-semiconductors</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/memory-is-now-redesigning-the-chip-hbm-fabs-long-term-deals-and-a-new-memory-architecture-this-week-in-semiconductors</guid><category><![CDATA[semiconductor industry news]]></category><category><![CDATA[memory]]></category><category><![CDATA[hbm]]></category><category><![CDATA[packaging]]></category><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Fri, 06 Nov 2026 04:02:36 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/9e367d4a-fc68-4fa6-b22b-341bc121ad44.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><img src="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/9e367d4a-fc68-4fa6-b22b-341bc121ad44.webp" alt="" /></p>
<p><em>Pradeep Kumar Adusumalli</em></p>
<p>Last week the industry's story was that memory is sold out for years. This week the industry answered — and the answer is reshaping chip design itself. A leading memory maker is reportedly building a new fab just to feed high-bandwidth memory demand, a major AI chip designer locked in multi-year memory supply, a robotics maker cut the memory on its next AI chips by a third to keep production moving, and a data-center chip designer validated a memory architecture built to sidestep high-bandwidth memory altogether. Four responses to one shortage — and every one lands on a physical design or STA engineer's desk.</p>
<h3>A new fab built around the memory stack</h3>
<p>The week's biggest construction story came from reports that a leading memory maker will build advanced foundry lines on its oldest chip campus, under a task force code-named Kairos, dedicated to the base dies underneath high-bandwidth memory stacks. Industry sources told one financial daily that the site will host 2-nanometer and 4-nanometer lines, with construction already under way, at an estimated investment of around 25 trillion won (about $19 billion); a second report put the estimate in the same range. (<a href="https://en.sedaily.com/finance/2026/11/04/samsung-to-build-advanced-foundry-fab-at-giheung-campus">Seoul Economic Daily</a>, <a href="https://voltgridnews.com/story/samsung-reportedly-plans-25-trillion-won-2nm-4nm-fab-at-giheung">VoltGrid News</a>)</p>
<p>The logic is a packaging story disguised as a fab story. A memory stack is a pile of DRAM core dies on a logic base die, and the base die is built on a foundry logic process. The reports say the maker already produces sixth-generation core dies at two other campuses, but the 4-nanometer base dies come from only some lines there. The new fab balances that out, and the 2-nanometer line is aimed at the next generation, whose base die is expected to move to the finer node. The same reports expect the maker's memory supply, in bits, to grow about 52 percent a year on average through 2028. None of this is confirmed by the maker: the reports cite industry sources and the investment figure is an estimate. For timing closure, the point is that the base die becomes a design variable — different stack heights, base-die nodes, and PHY configurations for the same accelerator — so the interface you sign off will have more variants, not fewer.</p>
<h3>Long-term deals replace spot buying</h3>
<p>The clearest confirmation that the shortage is now a planning assumption came on a major AI chip designer's third-quarter earnings call on November 3. Its chief executive said the company has long-term contracts with some memory suppliers and is discussing sharing higher component costs with customers — memory pricing has moved from a spot-market problem to a contract-structure problem. A second outlet's account of the same call reports the same quotes. (<a href="https://sammyguru.com/samsung-could-be-among-amd-key-long-term-memory-partners/">SammyGuru</a>, <a href="https://global.newspim.com/news/view/20261105000478">Newspim</a>)</p>
<p>The numbers explain it. One industry report projects contract pricing for sixth-generation high-bandwidth memory jumping about 121 percent year over year in 2027. One memory maker, on its own earnings call, said it had already sold out all of next year's planned memory supply, with sixth-generation sales more than tripling quarter over quarter to about 60 percent of its memory revenue. (<a href="https://tech-insider.org/hbm4-memory-shortage-amd-samsung-sk-hynix-price-spike-2026/">Tech Insider</a>, <a href="https://global.newspim.com/news/view/20261105000478">Newspim</a>)</p>
<p>The design consequence: when memory is allocated by contract, system architects design to the allocation. Build your closure flow so a stack-height change is a re-run, not a redesign — the memory interface on your next design may be specified late and changed mid-project.</p>
<h3>A robotics maker cuts memory to keep building</h3>
<p>The most direct evidence that the shortage is changing silicon — not just schedules — is a story first announced in early October that came back into focus this week. An electric-vehicle and robotics maker cut the planned memory on its next two AI chips: the nearer-term chip from 144 GB down to 96 GB of low-power DRAM, and the following generation from 216 GB down to 144 GB, a one-third cut. Its chief executive said on social media that it was "the only way to get enough volume" for its humanoid robot programme and that it "greatly reduces cost", arguing that memory bandwidth matters more than total capacity and that bandwidth would be kept as planned. (<a href="https://sammyguru.com/tesla-lowers-ai6-chip-memory-requirements-for-its-optimus-humanoid-robot/">SammyGuru</a>, <a href="https://tech-insider.org/tesla-ai5-chip-memory-cut-96gb-2026/">Tech Insider</a>)</p>
<p>A leading memory supplier has publicly estimated that a general-purpose humanoid robot will need more than 200 GB of DRAM — more than even the original, pre-cut specification — and the claim that the cut will have a negligible performance effect has no independent benchmark behind it yet. (<a href="https://www.latestevnews.com/articles/tesla-ai5-why-optimus-is-expected-to-make-do-with-96-gb">Latest EV News</a>, <a href="https://optimusk.blog/blog/tesla-optimus-hardware-specs/">Optimus Blog</a>)</p>
<p>The engineering fact stands regardless: a chip's memory capacity is now a supply-chain decision, not just an architecture decision. Fewer DRAM packages around the compute die changes the package, the power delivery, and the thermal map — a redesign of your problem if you work on package-level timing or power integrity, not just theirs.</p>
<h3>An alternative memory architecture reaches silicon</h3>
<p>On November 2, a data-center chip designer announced that its first-generation High Bandwidth Compute silicon has passed lab validation: the full memory read-write path and near-memory computation work on real hardware. The architecture stacks low-power DRAM on top of an AI accelerator die and computes close to the memory, instead of stacking DRAM beside a compute die on an interposer the way high-bandwidth memory does. (<a href="https://futurumgroup.com/insights/qualcomm-validates-first-hbc-silicon-with-an-18x-memory-bandwidth-gain-over-lpddr5x/">Futurum Group</a>, <a href="https://xenospectrum.com/en/qualcomm-hbc-silicon-validation/">XenoSpectrum</a>)</p>
<p>The headline figures are striking — 18 times the memory bandwidth of external low-power DRAM, and more than five times the bandwidth per watt of projected sixth-generation high-bandwidth memory specifications — but read them carefully. They are the designer's own figures, measured on a two-layer validation sample and normalized to target operating conditions, not an independent benchmark and not a finished product. The production version will stack four DRAM layers, and the first accelerator built on it is planned to sample in the first half of 2027. (<a href="https://superpowerdaily.com/posts/qualcomm-says-its-near-memory-ai-silicon-passed-a-key-lab-test">Superpower Daily</a>, <a href="https://nand-research.com/qualcomm-validates-hbc-silicon/">NAND Research</a>)</p>
<p>The idea matters more than the numbers: if memory is the bottleneck, make more of it, or move less data across it. This architecture bets on the second. For physical design, near-memory compute means the hottest, most timing-critical logic sits underneath a memory stack — a thermal and power-delivery problem unlike a conventional floorplan. If this ships at volume, a whole class of 3D thermal and IR-drop problems moves from research papers into signoff checklists.</p>
<h3>In brief: the mature-node exit accelerates</h3>
<p>A leading foundry has reportedly told customers it will exit its 8-inch wafer foundry business entirely by the end of 2030, redirecting investment toward 12-inch advanced processes, and has stopped taking new 8-inch product orders since the start of this year. The original report cites industry sources; a market research firm's analysis notes the 8-inch lines run mature processes from 65 to 180 nanometers — power management chips, display drivers, microcontrollers, image sensors — and projects average utilization across the top ten foundries rising toward 90 percent in 2026 as capacity shrinks. The same foundry is, in the same week, reportedly building a new advanced fab for memory base dies: capacity is being reallocated from mature nodes to the memory bottleneck, so engineers on mature-node products should expect tighter capacity and longer lead times, not relief. (<a href="https://www.thelec.net/news/articleView.html?idxno=15179">The Elec</a>, <a href="https://www.trendforce.com/news/2026/11/03/news-samsung-reportedly-to-exit-8-inch-foundry-by-2030-as-tsmc-also-cuts-capacity/">TrendForce</a>)</p>
<h3>Why this week matters if you close timing</h3>
<p><strong>First, memory configuration is now a variable, not a constant.</strong> Fab capacity, long-term contracts, and outright memory cuts all point the same way: the memory on your next design may change after the architecture is frozen. Parameterize the memory interface early.</p>
<p><strong>Second, the base die is a design object now.</strong> A fab built for base dies and an architecture that computes inside the memory stack both mean the logic under and around the memory is where the timing risk lives. Power delivery, thermal, and signal integrity at the memory interface deserve the same signoff attention as the core.</p>
<p><strong>Third, distrust any single number this week.</strong> The bandwidth claims are the vendor's own, the fab investment is an estimate from unnamed sources, and the memory-cut performance claim has a public rebuttal from a supplier. Every story here is real, and every one needs its caveat attached. The timing report you sign is only as good as the assumptions you checked.</p>
<p>This week the industry did not announce a faster transistor. It announced that memory is scarce enough to redesign fabs, rewrite contracts, shrink chip specifications, and invent a new memory architecture. All four will show up in your next timing report.</p>
<p><strong>Sources:</strong></p>
<p>Seoul Economic Daily — new fab for memory base dies (reported Nov 4): <a href="https://en.sedaily.com/finance/2026/11/04/samsung-to-build-advanced-foundry-fab-at-giheung-campus">https://en.sedaily.com/finance/2026/11/04/samsung-to-build-advanced-foundry-fab-at-giheung-campus</a>
VoltGrid News — fab investment estimate: <a href="https://voltgridnews.com/story/samsung-reportedly-plans-25-trillion-won-2nm-4nm-fab-at-giheung">https://voltgridnews.com/story/samsung-reportedly-plans-25-trillion-won-2nm-4nm-fab-at-giheung</a>
SammyGuru — long-term memory agreements (Nov 4): <a href="https://sammyguru.com/samsung-could-be-among-amd-key-long-term-memory-partners/">https://sammyguru.com/samsung-could-be-among-amd-key-long-term-memory-partners/</a>
Newspim — earnings call, cost sharing, memory prices (Nov 5): <a href="https://global.newspim.com/news/view/20261105000478">https://global.newspim.com/news/view/20261105000478</a>
Tech Insider — HBM4 price projection, supply picture: <a href="https://tech-insider.org/hbm4-memory-shortage-amd-samsung-sk-hynix-price-spike-2026/">https://tech-insider.org/hbm4-memory-shortage-amd-samsung-sk-hynix-price-spike-2026/</a>
SammyGuru — memory cut for robot production (Nov 5): <a href="https://sammyguru.com/tesla-lowers-ai6-chip-memory-requirements-for-its-optimus-humanoid-robot/">https://sammyguru.com/tesla-lowers-ai6-chip-memory-requirements-for-its-optimus-humanoid-robot/</a>
Tech Insider — memory cut timeline and quotes: <a href="https://tech-insider.org/tesla-ai5-chip-memory-cut-96gb-2026/">https://tech-insider.org/tesla-ai5-chip-memory-cut-96gb-2026/</a>
Latest EV News — memory figures and bandwidth argument: <a href="https://www.latestevnews.com/articles/tesla-ai5-why-optimus-is-expected-to-make-do-with-96-gb">https://www.latestevnews.com/articles/tesla-ai5-why-optimus-is-expected-to-make-do-with-96-gb</a>
Optimus Blog — memory cut and supplier dispute: <a href="https://optimusk.blog/blog/tesla-optimus-hardware-specs/">https://optimusk.blog/blog/tesla-optimus-hardware-specs/</a>
Futurum Group — HBC silicon validation: <a href="https://futurumgroup.com/insights/qualcomm-validates-first-hbc-silicon-with-an-18x-memory-bandwidth-gain-over-lpddr5x/">https://futurumgroup.com/insights/qualcomm-validates-first-hbc-silicon-with-an-18x-memory-bandwidth-gain-over-lpddr5x/</a>
XenoSpectrum — HBC validation details: <a href="https://xenospectrum.com/en/qualcomm-hbc-silicon-validation/">https://xenospectrum.com/en/qualcomm-hbc-silicon-validation/</a>
Superpower Daily — HBC lab test report: <a href="https://superpowerdaily.com/posts/qualcomm-says-its-near-memory-ai-silicon-passed-a-key-lab-test">https://superpowerdaily.com/posts/qualcomm-says-its-near-memory-ai-silicon-passed-a-key-lab-test</a>
NAND Research — HBC validation and roadmap: <a href="https://nand-research.com/qualcomm-validates-hbc-silicon/">https://nand-research.com/qualcomm-validates-hbc-silicon/</a>
The Elec — 8-inch foundry exit report (Nov 2): <a href="https://www.thelec.net/news/articleView.html?idxno=15179">https://www.thelec.net/news/articleView.html?idxno=15179</a>
TrendForce — 8-inch exit and utilization outlook (Nov 3): <a href="https://www.trendforce.com/news/2026/11/03/news-samsung-reportedly-to-exit-8-inch-foundry-by-2030-as-tsmc-also-cuts-capacity/">https://www.trendforce.com/news/2026/11/03/news-samsung-reportedly-to-exit-8-inch-foundry-by-2030-as-tsmc-also-cuts-capacity/</a></p>
]]></content:encoded></item><item><title><![CDATA[Layout-Dependent Effects: Stress, Well Proximity, and Timing Variation]]></title><description><![CDATA[Pradeep Kumar Adusumalli Two transistors can be drawn identically, pass every design rule, and still behave differently in silicon. The difference is not a…]]></description><link>https://thetapeoutlog.hashnode.dev/layout-dependent-effects-stress-well-proximity-and-timing-variation</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/layout-dependent-effects-stress-well-proximity-and-timing-variation</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Thu, 05 Nov 2026 04:17:45 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/25163807-531d-4412-a402-7c1f79209711.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Pradeep Kumar Adusumalli</em></p>
<p>Two transistors can be drawn identically, pass every design rule, and still behave differently in silicon. The difference is not a mistake in the schematic. It is the layout around each device: how far its gate sits from the edge of its diffusion, how close it is to a well edge, and what its neighbours look like. These are layout-dependent effects, or LDE, and they are one of the reasons a design that simulates cleanly before layout can shift after it.</p>
<p>LDE matters because it is systematic, not random. Random variation scatters device behaviour in all directions and is handled with margins and statistics. LDE moves behaviour in a direction you can predict from the layout — if you know what to look for. Two matched devices with different surroundings will mismatch on purpose, in effect, even though nobody intended it. And in digital design, two instances of the same standard cell can have different delays depending on what is placed next to them.</p>
<h2>Why the surroundings change the device</h2>
<p>A transistor is not an isolated drawing. It is built in silicon, surrounded by isolation oxide, wells, and neighbouring devices, and each of those surroundings leaves a physical mark on it. Three mechanisms do most of the work:</p>
<ul>
<li><strong>Mechanical stress</strong> from shallow trench isolation (STI) and other strained layers changes carrier mobility and threshold voltage.</li>
<li><strong>Non-uniform doping</strong> near well edges changes threshold voltage.</li>
<li><strong>Neighbourhood geometry</strong> — the spacing of diffusions and polysilicon around a device — changes stress, lithography, and local processing conditions.</li>
</ul>
<p>These effects grow as geometries shrink. Published guidance puts their serious onset from around the 90 nm generation onward, and they have only become more prominent at each smaller node, including FinFET nodes, where the three-dimensional device structure adds its own edge and neighbourhood context.</p>
<h2>STI stress and the length of diffusion</h2>
<p>Between devices sits shallow trench isolation: oxide-filled trenches that separate one active area from the next. The oxide and the silicon expand and contract differently, so the trench puts mechanical stress into the silicon around it. Stress changes how easily carriers move through the channel — mobility — and it also shifts threshold voltage and saturation current.</p>
<p>The key layout parameter is the length of diffusion, or LOD: the distance from the gate to the edge of the diffusion on each side, usually called SA and SB. A device with a long diffusion sits far from the trench edges and feels less stress; a device with a short diffusion sits close to the trench and feels more. The effect is also asymmetric: stress that helps one device type hurts the other. Tensile stress improves electron mobility and therefore helps NMOS devices, while compressive stress improves hole mobility and helps PMOS devices. So the same short diffusion that speeds up a PMOS device can slow down an NMOS device, and vice versa.</p>
<p>This is why STI stress cannot be treated as a uniform corner shift. It depends on each device's own SA and SB, on orientation, and on the direction of the stress relative to the channel. Published examples put some layout-dependent shifts in the tens-of-percent range for device current, which is far too large to absorb silently into a general margin on a timing-critical path.</p>
<h2>The well proximity effect</h2>
<p>The second major effect comes from the well itself. Wells are formed by high-energy ion implantation through a photoresist mask. Near the mask edge, some ions scatter off the resist sidewall and land where they were not intended, so the doping concentration near a well edge is not the same as deep inside the well. A transistor placed close to the well edge therefore sits in differently doped silicon than an identical transistor placed in the middle of the well.</p>
<p>The result is a threshold voltage shift that depends on distance from the well edge. Published work describes shifts of several to tens of millivolts for devices near the edge, with larger shifts reported for edge devices in some studies, and the effect can extend a micrometre or more from the well boundary. For a current mirror or a differential pair, that is a mismatch built directly into the layout. For a digital path, it is a delay shift on whichever device happens to sit near the edge.</p>
<p>In device models, the well proximity effect is captured with the SCA, SCB, and SCC parameters, which summarise the device's distances to the four well edges, together with fitting parameters that scale the resulting threshold voltage, mobility, and body-effect shifts. The important point for a layout engineer is the trend, not the formula: the closer the device is to the well edge, the larger the parameters, and the larger the shift.</p>
<h2>The rest of the neighbourhood: OSE, PSE, and dummy effects</h2>
<p>STI stress and well proximity get most of the attention, but they are not the whole story. Two more spacing effects show up regularly in advanced-node models:</p>
<ul>
<li><strong>Oxide-to-oxide spacing effect (OSE):</strong> how far a device's active area sits from neighbouring active areas changes the STI stress it feels. Two devices with the same SA and SB can still differ if one is surrounded by dense active areas and the other sits in open space.</li>
<li><strong>Poly spacing effect (PSE):</strong> the distance from a device's gate to neighbouring polysilicon shapes affects its electrical behaviour, through lithography and local processing conditions. Matched devices need matched poly spacing, not just matched gate length.</li>
</ul>
<p>Guard rings and dummy fill add a further layer. A guard ring changes the diffusion environment around the devices it protects, and dummy diffusion, poly, and metal fill change local density and stress. Dummy structures are still worth using — they give critical devices a uniform neighbourhood — but they must be placed consistently, because an inconsistent dummy environment is itself a source of mismatch. The rule is uniformity: whatever surrounds one matched device should surround the others in the same way.</p>
<h2>How models capture LDE, and what to check</h2>
<p>Modern compact models, starting with BSIM4, include LDE parameters for exactly this reason: SA and SB for STI stress, and SCA, SCB, and SCC for well proximity, along with parameters that scale their effect on mobility, threshold voltage, and body effect. Foundry process design kits supply the fitted values for a given process, and extraction tools compute the layout parameters from the finished layout. That means LDE-aware simulation is a post-layout activity by nature: the schematic cannot know SA, SB, or well-edge distances that do not exist yet.</p>
<p>For digital physical design, the practical consequence is that standard-cell timing is context-dependent. A cell characterised in isolation, or in one abutment context, does not have exactly the same delay when placed next to different neighbours, flipped, or shifted by a row. Research on detailed placement has shown that specific cell abutments can cause timing degradation on critical paths through exactly these effects. So LDE-aware timing analysis after placement — and re-checking after late placement changes or ECOs — is part of closing timing honestly, not an optional refinement.</p>
<h2>What PD engineers should do about it</h2>
<p>Most LDE discipline is layout discipline, and it costs little if it is applied early:</p>
<p><strong>1. Match the context, not just the device.</strong> For matched pairs, mirrors, and other sensitive structures, use the same diffusion size and shape, the same orientation, the same poly spacing, and the same distance to well edges. Common-centroid and interdigitated layouts exist largely to enforce this.</p>
<p><strong>2. Keep critical devices away from well edges.</strong> If a device must sit near a well edge, make sure its matched partners sit at the same distance, or accept and model the resulting shift rather than discovering it in silicon.</p>
<p><strong>3. Surround critical devices with dummies.</strong> Dummy devices at the edges of a matched group give every real device the same neighbourhood. The dummies absorb the edge effects so the real devices do not have to.</p>
<p><strong>4. Treat guard rings as part of the environment.</strong> A guard ring protects against latch-up and noise, but it also changes the diffusion context. Keep guard-ring placement symmetric around matched structures.</p>
<p><strong>5. Run LDE-aware checks on critical paths.</strong> After placement, check whether the cells on timing-critical paths sit in unusual contexts — short diffusions, well-edge proximity, or asymmetric neighbours — and confirm the timing analysis used models that include those effects. A path that closes in a context-free analysis and fails in an LDE-aware one is telling you something real.</p>
<h2>The takeaway</h2>
<p>Layout-dependent effects are the layout talking back to the circuit. STI stress shifts mobility through the length of diffusion, well proximity shifts threshold voltage through distance to the well edge, and oxide and poly spacing add further neighbourhood-dependent shifts. None of this is random noise to be margined away blindly: it is systematic, it is predictable from geometry, and it is modelled in the device parameters your tools already carry. The PD engineer's job is to make the layout boring in the right way — same context for matched devices, dummies at the edges, critical devices away from well boundaries — and to verify timing with models that know what the layout actually looks like. Do that, and the device you simulated is the device you built.</p>
<h2>Sources</h2>
<ul>
<li>Silvaco — Well Proximity and STI Stress (application note): <a href="https://Silvaco.com/wp-content/uploads/content/appNotes/iccad/2-002_StressEffectWellProximity.pdf">https://Silvaco.com/wp-content/uploads/content/appNotes/iccad/2-002_StressEffectWellProximity.pdf</a></li>
<li>EE News Europe — Layout-dependent effects in analog design: <a href="https://www.eenewseurope.com/en/layout-dependent-effects-in-analog-design/">https://www.eenewseurope.com/en/layout-dependent-effects-in-analog-design/</a></li>
<li>EDN — Variability in Small Geometries Can Cause Unexpected Problems: <a href="https://www.edn.com/variability-in-small-geometries-can-cause-unexpected-problems/">https://www.edn.com/variability-in-small-geometries-can-cause-unexpected-problems/</a></li>
<li>EDN — IC Performance: Challenges in Layout with CMOS Technology Nodes: <a href="https://www.edn.com/ic-performance-challenges-in-layout-with-cmos-technology-nodes/">https://www.edn.com/ic-performance-challenges-in-layout-with-cmos-technology-nodes/</a></li>
<li>AnalogHub — Layout-Dependent Effects: <a href="https://analoghub.ie/category/Layout/article/layoutDependentEffects">https://analoghub.ie/category/Layout/article/layoutDependentEffects</a></li>
<li>DATE 2023 — Mitigating Layout Dependent Effect-Induced Timing Risk in Multi-Row-Height Detailed Placement: <a href="https://past.date-conference.com/proceedings-archive/2023/DATA/318.pdf">https://past.date-conference.com/proceedings-archive/2023/DATA/318.pdf</a></li>
<li>University of British Columbia — BSIM4 DC Model (well-proximity and stress effect modelling): <a href="https://people.ece.ubc.ca/robertor/Links_files/Files/ICCAP-2008-doc/icmdl/icmdl044.html">https://people.ece.ubc.ca/robertor/Links_files/Files/ICCAP-2008-doc/icmdl/icmdl044.html</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[Parasitic Extraction: Why Your Post-Route Timing Moved]]></title><description><![CDATA[Pradeep Kumar Adusumalli The design closes timing after placement and clock tree synthesis. Routing finishes. Then the post-route report comes back with…]]></description><link>https://thetapeoutlog.hashnode.dev/parasitic-extraction-why-your-post-route-timing-moved</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/parasitic-extraction-why-your-post-route-timing-moved</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Tue, 03 Nov 2026 14:11:36 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/244556dc-81ae-4a31-b121-8f13b0f721b9.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Pradeep Kumar Adusumalli</em></p>
<p>The design closes timing after placement and clock tree synthesis. Routing finishes. Then the post-route report comes back with new violations on paths that were clean a day earlier. Nothing in the logic changed. What changed is that the wires stopped being estimates and became real.</p>
<p>That is what parasitic extraction is about, and learning to read its output is one of the most practical skills in physical design.</p>
<h2>What extraction actually measures</h2>
<p>A wire is not an ideal connection. It is a thin strip of metal with resistance, sitting close to other wires, to the substrate, and to metal layers above and below it. Every routed net therefore carries two kinds of unwanted baggage:</p>
<ul>
<li><strong>Resistance (R):</strong> set by the metal's resistivity, the wire's length, and its cross-section. Longer wires and thinner lower metals have more resistance. Every via adds resistance too, so a net that changes layers many times pays for each jump.</li>
<li><strong>Capacitance (C):</strong> every wire stores charge against its surroundings — partly to ground and the power rails, and partly to neighbouring wires. The second part causes the most surprises.</li>
</ul>
<p>Before routing, the timing tool cannot know any of this precisely, because the wires do not exist yet. After routing, an extraction tool reads the finished layout, layer by layer, and computes each net segment's R and C from the process data for metal thickness, spacing, and dielectric properties — a distributed RC network per net, not a single lumped number.</p>
<h2>SPEF: the file that carries the answer</h2>
<p>Extraction results are usually handed to the timing tool in SPEF, the Standard Parasitic Exchange Format, documented in IEEE 1481-1999. A SPEF file has a header (design name, extraction tool, and — critically — the units for time, capacitance, resistance, and inductance, which vary between tools), a name map that shortens long net names, a port section, and one detailed net section per net, marked <code>*D_NET</code>. Inside each block:</p>
<ul>
<li><strong><code>*CONN</code>:</strong> the pins the net connects — driver output and load inputs.</li>
<li><strong><code>*CAP</code>:</strong> the capacitance network. A line with one node is capacitance to ground; a line with two nodes is coupling capacitance between two nets. A net appearing in another net's coupling entries is its neighbour on the layout.</li>
<li><strong><code>*RES</code>:</strong> the resistance network, including via resistance where the net changes layers.</li>
</ul>
<p>The total capacitance on the <code>*D_NET</code> line is the sum of ground and coupling capacitance. For a slow net, comparing coupling against ground entries tells you at once whether the problem is a long resistive wire or a crowded neighbour.</p>
<h2>From estimate to extracted</h2>
<p>Timing accuracy improves in steps as the layout becomes more real:</p>
<ol>
<li><strong>Synthesis — wireload models.</strong> With no placement, wire length is estimated from statistics: fanout and block size, via tables mapping fanout to an expected length, multiplied by per-unit-length R and C. Nothing is known about cell positions or router layer choices.</li>
<li><strong>After placement — Steiner estimates.</strong> Cell coordinates allow an estimated route shape with per-layer RC values — but no congestion detours, via stacks, or future neighbours.</li>
<li><strong>After global routing — grid-level estimates.</strong> Nets are assigned to routing regions and layers, so layer choices are reflected, but the coarse grid misses short detours and exact neighbour spacing.</li>
<li><strong>After detailed routing — extraction.</strong> The real wires exist: actual length, layer, width, via count, neighbour spacing, and coupling to every adjacent net. This is the timing signoff is based on.</li>
</ol>
<p>Each step can move slack in either direction. A net the wireload model overestimated may come back faster; a net that looked short on paper may run a long way around a congested block on a thin, resistive layer. That is normal — and why post-route timing is a separate signoff step, not a formality.</p>
<h2>RC corners: a second axis on top of PVT corners</h2>
<p>Engineers already think in PVT corners — slow, typical, and fast combinations of process, voltage, and temperature, each with its own cell library. Extraction adds its own axis: the R and C values themselves are scaled for best- and worst-case process conditions, producing separate SPEF files:</p>
<ul>
<li><strong>Maximum RC (slow):</strong> high R, high C — for setup analysis, where the slowest paths matter.</li>
<li><strong>Minimum RC (fast):</strong> low R, low C — for hold analysis, where the fastest paths matter.</li>
<li><strong>Nominal RC:</strong> typical values, for general verification.</li>
</ul>
<p>Pair them correctly: setup uses the slow library corner with maximum parasitics, hold uses the fast corner with minimum parasitics. Mix them up and you quietly remove the pessimism hold signoff is supposed to carry — a path the report calls fine can then fail in silicon.</p>
<h2>Coupling capacitance: why neighbours change your delay</h2>
<p>Ground capacitance slows a net predictably: more capacitance, more delay. Coupling capacitance is trickier, because its effect depends on what the neighbouring net — the aggressor — is doing when the victim switches. If the aggressor switches in the <strong>opposite direction</strong>, the effective capacitance the victim driver sees grows larger than the coupling capacitance itself, and the victim slows down. If it switches in the <strong>same direction</strong>, the effective capacitance shrinks, and the victim speeds up. If it is quiet, the coupling capacitance behaves like ordinary ground capacitance.</p>
<p>So timing on the same extracted data can shift as analysis matures: the tool must decide which aggressors could plausibly switch at the same time (their timing windows). Early simplifications that treat all coupling as grounded capacitance are pessimistic for some paths and optimistic for others, so paths margined by pessimism can newly appear as violations once timing windows firm up.</p>
<p>Coupling also matters for signal integrity — a quiet victim can pick up a noise bump from a fast-switching aggressor — but delay change alone can break setup or hold. Long parallel runs on minimum-spacing lower metals are where coupling dominates; in dense layouts it can be most of a net's total capacitance.</p>
<h2>How to read the difference when timing moves</h2>
<p>Do not fix violations in the order the tool prints them. Read the change first:</p>
<p><strong>1. Compare stage to stage, per endpoint.</strong> Diff the post-route report against the post-CTS report for the same endpoints. Endpoints that only just went negative are usually one net away from closing; endpoints that moved a lot point to a structural problem — a rerouted net, a layer change, or a new coupling neighbour.</p>
<p><strong>2. Split net delay from cell delay.</strong> Cell-delay-dominated paths need sizing or logic changes; net-delay-dominated paths need a look at the net itself — length, layer, via count, and total capacitance in the SPEF.</p>
<p><strong>3. Check the coupling share.</strong> Compare coupling against ground capacitance in the SPEF <code>*CAP</code> section. If coupling dominates, the neighbours matter more than the wire itself: check what runs alongside it and whether the violation survives tighter timing windows.</p>
<p><strong>4. Check the corner pairing.</strong> Before touching the layout, confirm setup used maximum RC with the slow corner and hold used minimum RC with the fast corner. A mix-up produces violations no buffering will fix cleanly, because the underlying numbers are wrong.</p>
<p><strong>5. Look at layer and via choices.</strong> A net routed mostly on thin lower metals with many vias is resistive even if short. Promoting a long critical net to an upper, thicker layer, or cutting its via count, can recover more delay than adding buffers — at less power and area cost.</p>
<h2>What to do, once you know the cause</h2>
<p>A resistive net wants a thicker upper layer, a shorter route, or a repeater to split the RC delay (buffering helps long wires, not short nets loaded by capacitance). A capacitive load wants a bigger driver or a split net. A coupling-dominated net wants more spacing, shielding with quiet ground or power rails, a rerouted aggressor, or a less crowded layer — shielding costs tracks, so reserve it for critical nets. And if the violation rests on aggressors that cannot realistically switch together, tighten the analysis but document the assumption: that is a signoff decision, not a silent edit.</p>
<p>One standing discipline: re-extract after every ECO. A fix that changes routing changes neighbours' coupling too, so a report based on a stale SPEF is fiction.</p>
<h2>The takeaway</h2>
<p>Your post-route timing did not move because the design got worse. It moved because the timing model got honest. Wireload models guess, placement estimates approximate, and only extraction measures — the real length, the real layers, the real vias, and the real neighbours. The job at this stage is not to be surprised by the difference, but to read it: which nets moved, whether resistance, ground capacitance, or coupling moved them, and whether the corner pairing is right. Answer those three questions from the report and the SPEF, and post-route closure stops being a mystery and becomes a checklist.</p>
<h2>Sources</h2>
<ul>
<li>ChipVerify — Parasitics After Routing: <a href="https://chipverify.com/physical-design/parasitics-after-routing">https://chipverify.com/physical-design/parasitics-after-routing</a></li>
<li>ChipVerify — Post-Route STA: What Changes from Synthesis: <a href="https://chipverify.com/physical-design/post-route-sta">https://chipverify.com/physical-design/post-route-sta</a></li>
<li>VLSI Expert — How To Read SPEF: <a href="http://www.vlsi-expert.com/2010/08/how-to-read-spef.html">http://www.vlsi-expert.com/2010/08/how-to-read-spef.html</a></li>
<li>SiliconVLSI — Overshoot and Undershoot Glitch in VLSI: 7 Essential Crosstalk Concepts: <a href="https://siliconvlsi.com/overshoot-and-undershoot-glitch/">https://siliconvlsi.com/overshoot-and-undershoot-glitch/</a></li>
<li>DeepWiki (Caravel) — Parasitic Extraction (SPEF): <a href="https://deepwiki.com/efabless/caravel/4.3-parasitic-extraction-(spef)">https://deepwiki.com/efabless/caravel/4.3-parasitic-extraction-(spef)</a></li>
<li>Mantra VLSI — Wire Load Model (WLM): <a href="http://mantravlsi.blogspot.com/2014/08/wire-load-model-wlm_1.html">http://mantravlsi.blogspot.com/2014/08/wire-load-model-wlm_1.html</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[Liberty Files Demystified: What PD Engineers Should Read in a .lib]]></title><description><![CDATA[By Pradeep Kumar Adusumalli Every timing report you have ever read is, at bottom, a reading of a Liberty file. Synthesis, placement, routing, and static…]]></description><link>https://thetapeoutlog.hashnode.dev/liberty-files-demystified-what-pd-engineers-should-read-in-a-lib</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/liberty-files-demystified-what-pd-engineers-should-read-in-a-lib</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Mon, 02 Nov 2026 06:59:01 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6ae34bbecabb70c219e89ef2/34944300-6ecf-4094-a123-e13a14a70da7.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>By Pradeep Kumar Adusumalli</em></p>
<p>Every timing report you have ever read is, at bottom, a reading of a Liberty file. Synthesis, placement, routing, and static timing analysis never simulate your transistors during a timing run. They look things up. The .lib file is the table they look things up in: for each standard cell, how long it takes to switch, how much power it burns, and what constraints it must obey, all measured in advance by simulating the cell across input slews and output loads. If you understand what is actually inside a .lib, most timing mysteries get much less mysterious.</p>
<h2>The file is a contract, not a schematic</h2>
<p>A Liberty file is a plain-text, human-readable description of a cell library, written in a format defined by the Liberty standard. One file generally covers one process, voltage, and temperature corner, which is why a library ships as a set of files rather than a single one: a slow corner, a typical corner, and a fast corner are different files, because the same cell genuinely behaves differently at each. When a timing report surprises you, the first question is often embarrassingly simple: which corner's file was loaded?</p>
<p>The file has two scopes. The library header holds information common to the whole library: the units for time, capacitance, voltage, and current; the voltage and temperature thresholds used to measure slew and delay; the default operating conditions; the lookup-table templates; and the delay model in use. Then comes one section per cell, each describing that cell's function, area, pins, timing arcs, power, and constraints. Everything a tool needs to time a design lives in those two scopes, and nothing else is consulted at STA time.</p>
<h2>How a delay is actually looked up</h2>
<p>Take a simple combinational arc, say from input A to output Z of a two-input AND gate. The .lib does not store a single delay number for that arc. It stores a two-dimensional table. One axis is the input transition time, or slew, arriving at A. The other axis is the capacitive load on Z: the input capacitances of everything Z drives, plus the wire capacitance. The tool finds where your actual slew and load fall in that table and interpolates between the surrounding entries.</p>
<p>Two details matter here. First, rise and fall are tabulated separately, because a cell rarely pulls up and pulls down with the same strength. A report that shows different rise and fall delays for the same arc is not inconsistent; it is reading two different tables. Second, the table only covers the range it was characterized over. If your real slew or load falls outside that range, the tool extrapolates, and extrapolated delays are where reports quietly go wrong. A slew violation or an enormous load is often worth fixing before arguing with the delay number itself.</p>
<p>A minimal arc looks like this in the file's structure:</p>
<pre><code>cell(AND2) {
  area : 2;
  pin(A) { direction : input; }
  pin(Z) {
    direction : output;
    function : "A*B";
    timing() {
      related_pin : "A";
      timing_type : "combinational";
      cell_rise(...) { /* delay table */ }
      cell_fall(...) { /* delay table */ }
    }
  }
}
</code></pre>
<p>That skeleton is the whole idea: a cell, a pin, a function, and a timing arc with tables. Everything else in the file is elaboration on it.</p>
<h2>NLDM: the voltage-source model</h2>
<p>The older and simpler delay model is NLDM, the non-linear delay model, also called the table-lookup model. It treats the driving cell as a voltage source behind a fixed resistance, and the load as a single capacitor. Delay and output slew come straight from the two-dimensional tables described above.</p>
<p>NLDM is fast to read, compact, and accurate enough for most nets. Its weakness is exactly its simplification. A real receiver pin does not present a fixed capacitance: as the input switches, the capacitance it shows the driver changes, partly because of the Miller effect through the gate-drain overlap. And a real driver does not behave like a fixed resistor when it faces a long, resistive wire. On nets where the wire resistance is large compared with the driver's own resistance, the voltage-source approximation breaks down, and NLDM delay can drift noticeably from circuit simulation. That drift is what pushed the industry to a second model.</p>
<h2>CCS: the current-source model</h2>
<p>CCS, the composite current source model, replaces the voltage source with a current waveform. Instead of one delay number per slew-and-load combination, the file stores the driver's output current sampled over time for each combination, plus a receiver model that describes how the input pin's capacitance varies with input slew and output load.</p>
<p>Why that helps: a current source can drive an arbitrary resistor-capacitor network, so the tool can compute the waveform at the far end of a real, resistive net rather than pretending the net is a lumped capacitor. And the two-capacitance receiver model captures how the pin's capacitance changes during the transition, which is what makes CCS accurate on cells with a large Miller effect. Vendor material reports CCS stage delay and slew typically within about two percent of transistor-level simulation; treat that as a vendor-reported figure, not an independent benchmark, but it explains why CCS became the signoff model as geometries shrank.</p>
<p>The price is data volume and runtime. Current waveforms sampled over time for every slew and load point make CCS libraries much larger than NLDM ones, and delay calculation is heavier. That is a real trade-off, and it is why NLDM still appears in flows where speed matters more than the last percent of accuracy. When you see both models for the same library, the choice is not about which is "correct" in the abstract. It is about how much accuracy the nets you are closing actually need.</p>
<h2>Power: three numbers, not one</h2>
<p>Power in a .lib comes in three parts, and confusing them is a classic beginner mistake. Internal power is the energy burned inside the cell when it switches, tabulated against input slew and output load just like delay. Switching power is the energy spent charging and discharging the load capacitance, which the tool computes from the load itself. Leakage power is the static drain when the cell is idle, and it depends on the input state: the same cell leaks differently with its inputs at 00, 01, 10, or 11, so the file carries leakage values per state.</p>
<p>For a PD engineer, the practical reading is this: if a power report is dominated by leakage, the lever is cell choice, particularly threshold-voltage mix, not activity. If it is dominated by internal and switching power, the lever is activity, load, and drive strength. The .lib tells you which of those stories you are in, cell by cell.</p>
<h2>The constraints section is where signoff lives</h2>
<p>For sequential cells, the .lib also carries the timing checks: setup and hold times for data pins, recovery and removal times for asynchronous controls like reset, and minimum pulse widths for clocks and control pulses. These are tabulated too, against the relevant slews, because a setup requirement at a slow input edge is not the same as at a fast one.</p>
<p>This is worth pausing on, because it reframes what STA is doing. The tool is not inventing setup and hold numbers. It is checking your design against numbers the library vendor measured for that exact flop, at that corner, with those slews. When a setup violation appears on a flop, the .lib is the first place to look for whether the check itself is being read at the right corner and the right slew, before assuming the data path is at fault.</p>
<h2>Drive strength and threshold voltage: the same function, many cells</h2>
<p>Open a library and you will find the same logic function repeated many times: a small inverter, a medium one, a large one, and each of those again in low, regular, and high threshold-voltage variants. These are not duplicates. A stronger drive cell has wider transistors, so it switches a heavy load faster but costs more area and power. A low-threshold cell switches faster but leaks more; a high-threshold cell leaks less but switches slower.</p>
<p>That menu is the PD engineer's real vocabulary. Upsizing a cell on a critical path, swapping a low-threshold cell for a high-threshold one on a non-critical path to save leakage, and balancing the two across a design are all choices made in terms of these library variants. The .lib is where you verify what each variant actually buys you, instead of guessing from the cell name.</p>
<h2>A checklist for reading a .lib when something looks wrong</h2>
<p>When a timing or power number does not make sense, work through the file in this order:</p>
<ol>
<li><strong>Corner.</strong> Is the loaded file the corner you think it is? A fast-corner file in a setup run, or vice versa, explains a surprising number of impossible-looking results.</li>
<li><strong>Model.</strong> Is the run using NLDM or CCS, and does that match what the signoff flow expects? A model mismatch between synthesis and STA is a quiet source of disagreement.</li>
<li><strong>Table range.</strong> Are the actual input slews and output loads inside the characterized range, or is the tool extrapolating? Fix slew and load violations before disputing the delay.</li>
<li><strong>Rise versus fall.</strong> Are you comparing the right tables? Asymmetric rise and fall delays are normal.</li>
<li><strong>Pin capacitance.</strong> Is the load built from the receiver pins' capacitances plus wire capacitance, or is something double-counted or missing? In CCS, remember the receiver capacitance itself varies with slew and load.</li>
<li><strong>Design rules.</strong> Check the file's maximum transition, maximum capacitance, and maximum fanout limits. A cell pushed past its own limits is not a library problem; it is a design problem the library is correctly flagging.</li>
<li><strong>Variation handling.</strong> The .lib gives you nominal corner data. On-chip variation is handled in the flow with corner libraries plus derating or advanced on-chip variation settings in the tool. If a report's margins look wrong, check those settings rather than hunting for a single derating number inside the file.</li>
</ol>
<h2>The habit that pays off</h2>
<p>You do not need to memorize a .lib. You need to know how to open one, find a cell, find an arc, and read its tables and constraints. Do that once for a cell on a real critical path, and again for a flop with a setup violation, and the file stops being a black box. From then on, every timing report reads differently: not as a verdict handed down by the tool, but as a lookup you can trace back to a table, a corner, and a model, and argue with on its own terms. That is the difference between debugging timing and merely re-running it.</p>
<h2>Sources</h2>
<ul>
<li>SiliconCompiler glossary: Liberty, NLDM, and CCS delay models — <a href="https://docs.siliconcompiler.com/en/latest/user_guide/glossary.html">https://docs.siliconcompiler.com/en/latest/user_guide/glossary.html</a></li>
<li>VLSI Talks: timing library (.lib) structure, NLDM and CCS — <a href="https://vlsitalks.com/physical-design/inputfiles/">https://vlsitalks.com/physical-design/inputfiles/</a></li>
<li>VLSI Master: stages of STA and the .lib file's contents — <a href="https://vlsimaster.com/stages-of-sta/">https://vlsimaster.com/stages-of-sta/</a></li>
<li>EE Times: CCS advanced delay calculation methodology — <a href="https://www.eetimes.com/ccs-offers-advanced-delay-calculation-methodology/">https://www.eetimes.com/ccs-offers-advanced-delay-calculation-methodology/</a></li>
<li>Liberty Reference Manual (R-2020.09) — <a href="https://zao111222333.github.io/liberty-db/2020.09/reference_manual.pdf">https://zao111222333.github.io/liberty-db/2020.09/reference_manual.pdf</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[Why STA Signoff Is Never Just a Checkbox]]></title><description><![CDATA[There is a particular silence that falls over a project room when the final signoff timing run comes back clean. Nobody cheers, exactly — everyone is too…]]></description><link>https://thetapeoutlog.hashnode.dev/why-sta-signoff-is-never-just-a-checkbox</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/why-sta-signoff-is-never-just-a-checkbox</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Thu, 29 Oct 2026 14:14:56 GMT</pubDate><content:encoded><![CDATA[<p>There is a particular silence that falls over a project room when the final signoff timing run comes back clean. Nobody cheers, exactly — everyone is too tired for that — but shoulders drop, and somebody says the words: timing is closed, we are done. I have learned to be suspicious of that silence. Not because a green report is bad news, but because a green report is a claim, and claims need evidence. Static timing analysis does not tell you the chip will work. It tells you the chip will work under the conditions, constraints, and libraries you gave the tool. Change any of those, and the color of the report changes with it. Signoff, done properly, is the discipline of making sure the green means something.</p>
<h2>A green report is only as good as its assumptions</h2>
<p>Every STA run answers a very specific question: given these libraries, these operating conditions, these constraints, and this extracted netlist — does every timing path meet its requirement? Notice how many givens that question contains. The tool is ruthlessly honest about the answer and completely silent about whether the question was the right one. That is why my first question about any clean report is never great, what is next — it is what did we assume? Which corners did we run? Are the constraints current? Do the exceptions still describe the real design? A report can be perfectly green and perfectly wrong if the assumptions underneath it are stale.</p>
<h2>Corners are not decoration</h2>
<p>The PVT mantra — process, voltage, temperature — is easy to recite and easy to under-run. Setup gets checked at the slow corner: slow process, low voltage, high temperature, where cells are at their laziest. Hold gets checked at the fast corner: fast process, high voltage, low temperature, where everything races. But real signoff is wider than the textbook. Multi-voltage designs need every domain's extremes. Low-power modes with power gating or retention add corners the default flow will not invent for you. Best-case and worst-case RC extraction corners interact with the PVT corners in ways that surprise people on every single tapeout. And then there are the corners you run because of a scar: the team that got burned by a min-delay failure at cold temperature never forgets the cold corner again. The corner list is a judgment call, and judgment calls are where signoff lives.</p>
<h2>Derates should be honest, not hopeful</h2>
<p>On-chip variation derating is where optimism goes to get quantified. A flat derate is a blunt instrument: simple, pessimistic in some places, optimistic in others. Depth- and distance-based derates sharpen the model; statistical approaches sharpen it further. Whatever method you use, the derate numbers have to be honest. I have watched teams shave derates to close timing at the end of a project. Understand what that transaction really is: you are trading analysis margin for schedule, and the silicon collects the debt. If the margin has to shrink, it should shrink because better modeling justified it — tighter libraries, real variation data — not because the calendar said so. A derate you shrink under pressure is not a derate anymore. It is a wish.</p>
<h2>Earn the green</h2>
<p>The constraints file is the most powerful and most dangerous document in the signoff flow. False paths, multicycle paths, case analysis, clock definitions — every one of them is a promise about how the design actually behaves, and the tool believes every word. Stale constraints are how green reports lie: a false path that was legitimate three RTL drops ago, a multicycle exception on a path whose protocol changed, a generated clock that no longer exists. I audit exceptions the way an accountant audits expenses, and every single one needs a reason that is still true today. Beyond that, my habits are simple: read the top of the violator list even on clean runs, ask why it is green and not just whether it is green, diff the assumptions and not just the results, and keep a signoff checklist that outlives any single project. The checkbox version of signoff asks: is the report green? The engineering version asks: do I understand why it is green — and would I bet silicon on that understanding? Only the second question is worth signing your name to.</p>
]]></content:encoded></item><item><title><![CDATA[CTS: Skew Isn't the Enemy, Unmanaged Skew Is]]></title><description><![CDATA[Early in my career, I believed the perfect clock tree was the one with zero skew. It sounded right: every flip-flop receiving its clock at exactly the same…]]></description><link>https://thetapeoutlog.hashnode.dev/cts-skew-isn-t-the-enemy-unmanaged-skew-is</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/cts-skew-isn-t-the-enemy-unmanaged-skew-is</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Thu, 29 Oct 2026 14:13:37 GMT</pubDate><content:encoded><![CDATA[<p>Early in my career, I believed the perfect clock tree was the one with zero skew. It sounded right: every flip-flop receiving its clock at exactly the same instant, a kind of digital utopia. Then a senior engineer showed me a design where the clock tree carried nearly a hundred picoseconds of deliberate, carefully managed skew — and it was closing timing that a perfect zero-skew tree could not. That was the day I learned the real lesson of clock tree synthesis: skew is not the enemy. Unmanaged skew is.</p>
<h2>What skew actually is</h2>
<p>Skew is simply the difference in clock arrival times between any two flip-flops. It exists because the clock must travel through buffers and wires to reach every corner of the die, and no two paths are identical. A surprising amount of engineering anxiety gets spent trying to drive this number to zero, as if zero were inherently virtuous. But zero skew is not a requirement — it is just one point in a large space of possibilities, and it is rarely the optimal one. What matters is not whether skew exists, but whether you know where it is, how much there is, and whether it is helping or hurting your timing. An engineer who can answer those three questions owns their clock tree. One who cannot is just hoping.</p>
<h2>Useful skew: borrowing time where it is needed</h2>
<p>Here is the insight that changed how I think about CTS: setup timing is about the relationship between the launching and capturing clocks, not about their absolute arrival times. If a critical path is failing setup by fifty picoseconds, you do not necessarily need a faster data path — you can delay the capturing clock by sixty picoseconds and the violation disappears. You have borrowed time from the next cycle. This is useful skew, and it is one of the most powerful timing-closure techniques in physical design. The catch is that borrowing must be repaid: delaying the capture clock eats into the hold margin of paths fed by that flip-flop, and it steals time from the following stage. Useful skew is a precision instrument, not a blunt one. Applied blindly, it creates as many violations as it fixes. Applied deliberately, with a clear picture of which paths need help and which have margin to lend, it can close timing that no amount of gate sizing ever will. The best CTS engineers I know spend more time studying timing reports than running the tool — because the report tells them where the borrowing should happen.</p>
<h2>Balanced versus biased trees</h2>
<p>A balanced tree tries to equalize insertion delay to all endpoints — the classic zero-skew goal. A biased tree deliberately shapes arrival times to favor critical paths. Neither is universally right. Balanced trees are predictable, easier to analyze, and the correct default for most designs. Biased trees are a timing-closure tool: when a handful of paths dominate your worst negative slack, shaping the clock to serve those paths can be the difference between closing and spinning another week. The key judgment is recognizing when the design has crossed from the tool can handle it into the clock needs intent. That judgment comes from reading timing reports, understanding where critical paths cluster physically and logically, and knowing when the clock is the right lever versus when the real problem is placement density, logic depth, or over-constrained exceptions. A biased tree built on a misunderstanding of the critical paths is just a slower, more confusing balanced tree.</p>
<h2>The engineering is in the asking</h2>
<p>Modern CTS tools are extraordinarily capable — but they are optimizers, not mind readers. They build exactly the tree you asked for, with the skew targets, latency targets, and useful-skew allowances you specified. Ask for zero skew everywhere and you will get it, along with the power and area it costs and the timing opportunities it forecloses. The real engineering in CTS happens before the tool runs: deciding the clock architecture, choosing which paths deserve intentional skew, setting latency budgets the design can afford, and defining what the power envelope allows. I have seen more CTS problems caused by thoughtless targets than by tool limitations — skew targets copied from a previous project, latency limits set without understanding the floorplan, useful skew disabled because someone feared it. Know what to ask for, verify what you got, and remember that the best clock tree is not the one with the smallest skew number. It is the one that helps the design meet timing with the least power and area. Skew is not the enemy. Not knowing what your skew is doing — that is the enemy.</p>
]]></content:encoded></item><item><title><![CDATA[IR Drop: The Silent Timing Killer]]></title><description><![CDATA[The most unsettling failure I have seen in this industry was not a timing violation. It was the absence of one. The STA report was clean — every corner…]]></description><link>https://thetapeoutlog.hashnode.dev/ir-drop-the-silent-timing-killer</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/ir-drop-the-silent-timing-killer</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Thu, 29 Oct 2026 14:12:22 GMT</pubDate><content:encoded><![CDATA[<p>The most unsettling failure I have seen in this industry was not a timing violation. It was the absence of one. The STA report was clean — every corner, every mode, green across the board. The chip came back from the fab and failed at speed. Not everywhere, not randomly, but in one specific region of the die, under one specific workload. The culprit was not in the timing report at all. It was in the voltage map. That was my introduction to IR drop: the silent timing killer.</p>
<h2>When STA is clean and silicon fails</h2>
<p>Static timing analysis assumes every cell sees its nominal supply voltage. Real silicon disagrees. As current flows through the resistive power delivery network, the voltage droops — sometimes by tens of millivolts, sometimes by far more. A cell operating at a reduced voltage is a slower cell, and a slower cell on a critical path is a setup violation that STA never saw, because STA was told the voltage was ideal. This is what makes IR drop so dangerous: it is a timing problem that hides from the timing tool. The standard libraries are characterized at fixed voltages, and unless your flow explicitly accounts for the droop, you are signing off a fantasy. Dynamic droop — the fast, workload-dependent kind — is the worst offender, because it strikes exactly when the chip is working hardest, which is exactly when timing matters most.</p>
<h2>Start rail analysis early, not at signoff</h2>
<p>The classic mistake is treating IR drop as a signoff checkbox, something to run once the design is nearly done. By then, the power grid is fixed, the floorplan is frozen, and any serious droop problem requires surgery. The right time for the first rail analysis is right after floorplanning, when the PDN is taking shape and changes are still cheap. Early analysis answers the structural questions: are the straps wide enough, are the vias sufficient, are the power pads placed to feed the hungry regions? I have watched teams discover at signoff that an entire block was starved because a strap was necked down to make room for a signal route six months earlier. That discovery costs weeks at signoff. It would have cost an afternoon in floorplan.</p>
<h2>Be honest about switching activity</h2>
<p>Rail analysis is only as good as the activity it assumes, and activity is where optimism sneaks in. Vectorless analysis with default toggle rates is a starting point, not an answer — it spreads switching evenly and misses the brutal reality that real workloads create hot spots. The honest approach uses real activity: simulated vectors from the verification team, realistic workload scenarios, the actual patterns the chip will run. I always ask for the worst realistic case, not the average case, because silicon does not grade on a curve. If the verification team cannot provide vectors, say so loudly and add margin — do not quietly accept a clean vectorless report as proof. A related trap is the clock tree: clocks toggle every cycle and burn real power, and an under-budgeted clock network will droop exactly when every flip-flop needs it most.</p>
<h2>Own the PDN after floorplan</h2>
<p>Here is an ownership problem I see repeatedly: the PDN gets designed during floorplan, and then nobody owns it again. Signal routing squeezes straps, ECOs punch holes in the grid, engineering change orders add hungry cells in already-marginal regions — and the rail analysis from three months ago quietly becomes fiction. Someone needs to own the power grid for the life of the project, re-validating it after every major change. That someone is usually the physical design engineer, because the PDN lives in the layout and its failures show up as timing. Make rail analysis part of your regular regression, not a one-time event. And when timing violations cluster in one region of the die — a corner, a block boundary, the far end from the power pads — check the voltage map before you touch the timing. I have seen engineers spend days upsizing gates in a droop region when the real fix was a wider strap two layers up. The voltage map does not lie, but you have to look at it.</p>
]]></content:encoded></item><item><title><![CDATA[DRC/LVS Clean Does Not Mean Tapeout Ready]]></title><description><![CDATA[There is a special kind of confidence that comes from a clean DRC/LVS report. The violation count reads zero, the layout matches the schematic, and it…]]></description><link>https://thetapeoutlog.hashnode.dev/drc-lvs-clean-does-not-mean-tapeout-ready</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/drc-lvs-clean-does-not-mean-tapeout-ready</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Thu, 29 Oct 2026 14:11:13 GMT</pubDate><content:encoded><![CDATA[<p>There is a special kind of confidence that comes from a clean DRC/LVS report. The violation count reads zero, the layout matches the schematic, and it feels like the hardest part is over. I have felt that confidence. I have also watched it evaporate in a tapeout review when someone asked about antenna checks — and the room went quiet. DRC and LVS are the foundation of physical verification, but they are not the building. Here is what I have learned about everything that has to happen after the zero-violation report.</p>
<h2>The checks beyond the checks</h2>
<p>Antenna rules exist because plasma etching can charge up long metal wires and damage the thin gate oxide they connect to. DRC decks check geometry; antenna checks verify that every gate has a safe discharge path, often requiring diode insertion or layer hopping. I have seen clean DRC runs fail antenna badly, especially in designs with long top-metal routes feeding small gates. Density rules are another quiet killer: chemical-mechanical polishing needs uniform metal density, and a region that is too sparse or too dense will polish unevenly, breaking planarity for the layers above. Fill insertion fixes it, but fill must be timed and placed correctly — dummy metal near sensitive analog or RF structures can couple noise where you least want it. These are not exotic concerns. They are standard signoff, and they fail more often than newcomers expect.</p>
<h2>Electromigration and ESD: the reliability duo</h2>
<p>Electromigration is the slow death of wires: over years of current flow, metal atoms migrate and voids form, until a wire that was perfect on day one opens on day one thousand. EM checks verify that current densities stay within foundry limits across every segment, under realistic switching activity. The trap is that EM is workload-dependent — a wire that is safe under average current can fail under peak current, and peak current happens in exactly the bursts your product is designed for. ESD structures are the complementary concern: the diodes and clamps that protect the chip from electrostatic discharge during handling and assembly. They must be present, correctly sized, correctly placed near the pads, and verified — not assumed. I treat EM and ESD as reliability signoff, distinct from geometric signoff, with their own checklists and their own reviewers. A chip that passes DRC but fails in the field from EM has not really passed anything.</p>
<h2>Seal ring and scribe-line compliance</h2>
<p>The seal ring is the moat around your castle: a continuous metal structure that protects the die from moisture ingress and scribe damage during wafer dicing. It has strict construction rules — width, continuity, via stacks, corner treatment — and it must be verified, not just drawn. Scribe-line compliance covers what happens at the die edge: keep-out regions, alignment marks, and the structures the foundry places between dice. These feel like afterthoughts until a die cracks during dicing or moisture creeps in during qualification. I have seen tapeouts delayed because the seal ring had a gap at one corner that nobody checked until the foundry's incoming QA flagged it. The die boundary deserves the same rigor as the core.</p>
<h2>The checklist after the checklist</h2>
<p>What ties all of this together is a simple practice: maintain a tapeout checklist that goes beyond DRC/LVS, and treat every first-time discovery as a process failure. The first time antenna bites you, the response should not just be to fix the violations — it should be to add antenna checks to the standard flow so it can never surprise you again. The same for density, EM, ESD, seal ring, and scribe line. Each painful discovery should harden into an automated check with a clear owner and a clear pass criterion. Over a career, this checklist becomes one of your most valuable assets: the distilled memory of everything that can go wrong after the obvious things go right. DRC/LVS clean is a milestone worth celebrating. Just do not confuse the milestone with the destination.</p>
]]></content:encoded></item><item><title><![CDATA[What a Timing ECO Taught Me About Working With the RTL Team]]></title><description><![CDATA[Late in a project, with the tapeout clock ticking, I was staring at a setup violation that would not die. The path ran from a deeply pipelined block…]]></description><link>https://thetapeoutlog.hashnode.dev/what-a-timing-eco-taught-me-about-working-with-the-rtl-team</link><guid isPermaLink="true">https://thetapeoutlog.hashnode.dev/what-a-timing-eco-taught-me-about-working-with-the-rtl-team</guid><dc:creator><![CDATA[Pradeep Kumar Adusumalli]]></dc:creator><pubDate>Thu, 29 Oct 2026 14:10:09 GMT</pubDate><content:encoded><![CDATA[<p>Late in a project, with the tapeout clock ticking, I was staring at a setup violation that would not die. The path ran from a deeply pipelined block through a congested region to a faraway endpoint, and every physical fix — upsizing, restructuring, useful skew — bought a little and cost a lot. The honest answer was that the path needed a pipeline stage, which meant an RTL change, which meant going to the RTL team three weeks before tapeout and asking them to touch working logic. I dreaded that conversation. It ended up teaching me more about engineering than any tool ever has.</p>
<h2>Bring options, not just violations</h2>
<p>My first instinct was to file the violation: here is the failing path, here is the slack, please fix. That is how I had seen it done, and it is also a terrible way to work. A bare violation report hands someone a problem with no context and an implicit demand, and it invites defensiveness. Instead, I did the analysis first. I framed the path three ways: what it would take to fix in physical design alone (painful, risky, power-hungry), what a single pipeline stage would buy (clean closure, small latency cost), and what a narrower micro-architectural tweak could do (a compromise neither side loved). I walked into the conversation with a one-page summary: the problem, the options, the trade-offs, and my recommendation. The RTL lead did not feel ambushed; they felt consulted. We picked the pipeline stage, the ECO went in cleanly, and timing closed. The lesson stuck: never report a problem without doing the homework to frame the solution space. Anyone can point at red text. An engineer brings options.</p>
<h2>Physical design lives at the intersection</h2>
<p>That ECO taught me something structural about our discipline. Physical design sits at the intersection of architecture, RTL, DFT, and signoff — we are the ones who discover, late and concretely, the consequences of decisions made months earlier in abstract. A pipeline depth chosen for frequency targets, a DFT hook that adds muxing to a critical path, a clock-gating strategy that complicates the tree — these all land on the PD engineer's desk as timing violations with no obvious owner. The engineers who thrive here are not just tool experts; they are translators. They can explain a physical problem in RTL terms, an RTL constraint in physical terms, and a schedule trade-off in terms both sides understand. If you only speak the language of your own stage in the flow, you will spend your career filing tickets. Learn enough about your neighbors' worlds to have real conversations, and the whole project moves faster.</p>
<h2>Soft skills are PD skills</h2>
<p>There is a persistent myth in backend engineering that communication is someone else's job — that the work is the layout, the timing, the reports, and that dealing with people is overhead. My experience is the opposite: the soft skills are PD skills. Timing closure is a team sport played under deadline pressure, and the engineer who can write a clear summary, run a focused debug meeting, disagree without burning bridges, and ask for help early will close more designs than the brilliant loner every time. The timing ECO story is really a communication story. The technical analysis mattered, but what made it work was empathy for the RTL team's position, clarity about what I needed and why, and respect for their judgment on their own logic. If you are starting in physical design, invest in these skills as deliberately as you invest in learning the tools. The tools change every few years. The ability to work with people compounds forever.</p>
]]></content:encoded></item></channel></rss>