<?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" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Sktbttr on Medium]]></title>
        <description><![CDATA[Stories by Sktbttr on Medium]]></description>
        <link>https://medium.com/@sktbttr?source=rss-50c86b95ce78------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*3ntpTUMx2DCJMXLZ</url>
            <title>Stories by Sktbttr on Medium</title>
            <link>https://medium.com/@sktbttr?source=rss-50c86b95ce78------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 04:13:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sktbttr/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Buffers, Inverters, or Both? Understanding Cell Selection in Clock Tree Synthesis]]></title>
            <link>https://medium.com/@sktbttr/buffers-inverters-or-both-understanding-cell-selection-in-clock-tree-synthesis-db163831b12b?source=rss-50c86b95ce78------2</link>
            <guid isPermaLink="false">https://medium.com/p/db163831b12b</guid>
            <dc:creator><![CDATA[Sktbttr]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 18:14:30 GMT</pubDate>
            <atom:updated>2026-10-29T18:14:30.063Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>Author: Shaik Thouheed<br>Email : sktbttr@gmail.com</em></strong></p><p><strong>Clock Tree Synthesis (CTS)</strong> is one of the most interesting stages of Physical Design.</p><p>At first glance, the objective looks simple:</p><blockquote><strong><em>Distribute the clock from the source to thousands of sequential elements while controlling skew, insertion delay, transition, and other electrical constraints.</em></strong></blockquote><p>But there is an interesting question hidden inside the clock tree:</p><p><strong>What cells should we allow CTS to use?</strong></p><p>For example:</p><ol><li>Only clock buffers</li><li>Only clock inverters</li><li>A combination of clock buffers and clock inverters</li></ol><p>Would all three approaches work?</p><p>Potentially, yes. But they don’t give the CTS engine the same optimization freedom.</p><p>Let’s understand why.</p><h3>First, What Is the Difference Between a Buffer and an Inverter?</h3><p>Logically, the difference is very simple.</p><h3>Clock Buffer</h3><p>A buffer maintains the same polarity. So if a rising edge enters the buffer, a rising edge eventually appears at its output.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/421/1*bXBelWi0mNq5ei_qq8d7sA.png" /></figure><h3>Clock Inverter</h3><p>An inverter changes the polarity. This creates an important consideration when we use inverters inside a clock tree. The final clock polarity reaching the sequential elements must remain correct.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/418/1*5JZ3LJYYEm3OhQj0oJ4qlQ.png" /></figure><h3>Case 1: CTS Using Only Clock Buffers</h3><p>This is conceptually straightforward because every stage preserves clock polarity. There is no inversion to keep track of.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OL1IVRlC0Er3zTpiaeFeWw.png" /></figure><h3>Advantages</h3><p>Using only clock buffers can provide:</p><ul><li>simple polarity management</li><li>straightforward clock-tree topology</li><li>easier debugging</li><li>predictable clock behavior</li><li>no requirement to balance inversion polarity</li></ul><p>The CTS engine can insert different buffer drive strengths depending on the fanout, wire capacitance, transition, and load.</p><h3>But There Is a Trade-Off</h3><p>A CMOS buffer can conceptually be thought of as being built from multiple inverter stages.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OJdKFxNudfbT99artMnRvw.png" /></figure><h3>Case 2: CTS Using Only Clock Inverters</h3><p>Now let’s make things more interesting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AR2qcAaw-WQbdVRTIEjtRw.png" /></figure><h3>Why Would We Want Clock Inverters at All?</h3><p>This is where the physical-design perspective becomes important.<br>An inverter is one of the fundamental CMOS gates.</p><p>For some library, load, slew, and drive-strength combinations, clock inverters may provide attractive characteristics in terms of:</p><ul><li>delay</li><li>area</li><li>power</li><li>capacitance</li><li>transition</li></ul><p>So completely excluding clock inverters may unnecessarily remove useful optimization choices from CTS.</p><p>However, there is another side to the story.</p><p>With inverter-based clock distribution, CTS must correctly manage clock polarity throughout the network.</p><p>So an inverter is not simply a drop-in replacement for a buffer everywhere.</p><h3>Case 3: Using Both Clock Buffers and Clock Inverters</h3><p>Now we give CTS both types of clock cells.</p><p>Now the CTS engine has a larger solution space.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CzmF3tLwKlg2Ny5VLu7OYw.png" /></figure><p>Instead of saying:</p><blockquote><em>“You must build my entire clock tree using buffers.”</em></blockquote><p>or:</p><blockquote><em>“You must build my entire clock tree using inverters.”</em></blockquote><p>we effectively say:</p><blockquote><em>“Here are the legal clock cells. Select appropriate cells while satisfying clock and electrical requirements.”</em></blockquote><p>This gives CTS more flexibility.</p><h3>What Is CTS Actually Trying to Optimize?</h3><p>The purpose of CTS isn’t simply to insert as few cells as possible.<br>It has multiple objectives and constraints to balance, including:</p><h3>1. Clock Skew</h3><p>We generally want controlled arrival-time differences between related sinks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8bsdOBr9vRucyE_92gOQxw.png" /></figure><h3>2. Insertion Delay</h3><p>Clock insertion delay is the time required for the clock to propagate through the clock network from the defined source/reference point toward the sink.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZNUO1Pxx2gMMvfCYp804dA.png" /></figure><h3>3. Transition</h3><p>Clock nets drive significant capacitance.If a clock edge becomes too slow:<br>it can create electrical and timing problems. Appropriate clock-cell sizing and placement help control slew.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_9BkfKHqYy9Nx87YWU6BAg.png" /></figure><h3>4. Fanout</h3><p>A single clock cell cannot practically drive unlimited loads.</p><p>CTS therefore creates hierarchy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fGtguluiF2sfQ0-QottfRw.png" /></figure><h3>5. Power</h3><p>Clock networks switch continuously whenever the clock is active.<br>That makes clock-tree power important.</p><p>A simple approximation for dynamic switching power is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SB701YEImEJPvQ7iKeI8cw.png" /></figure><h3>So Which Option Is Better?</h3><p>There is no universal answer that says:</p><blockquote><em>“Always use buffers.”</em></blockquote><p>or:</p><blockquote><em>“Always use inverters.”</em></blockquote><p>The choice depends on:</p><ul><li>standard-cell library</li><li>CTS methodology</li><li>technology node</li><li>PVT conditions</li><li>clock frequency</li><li>clock topology</li><li>duty-cycle requirements</li><li>power targets</li><li>transition constraints</li><li>skew targets</li><li>physical characteristics of the design</li></ul><p>However, <strong>allowing an appropriate combination of qualified clock buffers and clock inverters gives CTS a broader optimization space</strong> than artificially restricting it to only one cell type.</p><h3>One Important Misconception</h3><p>Sometimes we hear:</p><blockquote><em>“Clock inverters are always faster and consume less power than clock buffers.”</em></blockquote><p>That statement is too broad.</p><p>A more technically correct statement is:</p><blockquote><strong><em>Clock inverters can provide better power, area, capacitance, or delay characteristics for certain operating conditions, but the actual benefit depends on the characterized library cells, load, slew, drive strength, PVT corner, and physical implementation.</em></strong></blockquote><p>That distinction becomes especially important during real implementation and signoff.</p><h3>Interview Question: Why Do We Use Both Buffers and Inverters in CTS?</h3><p>If an interviewer asks:</p><blockquote><strong><em>Why are both clock buffers and clock inverters allowed during CTS?</em></strong></blockquote><p>I would answer:</p><blockquote><strong><em>Allowing both buffers and inverters gives the CTS engine greater flexibility to optimize clock skew, insertion delay, transition, power, area, and other electrical constraints. Inverters can be efficient for certain loads and drive conditions, while buffers simplify polarity handling. When inverters are used, CTS must ensure the required clock polarity is preserved at the sinks. Ultimately, the best choice depends on the characterized clock cells and the CTS methodology.</em></strong></blockquote><h3>Final Thoughts</h3><p>One thing I have learned while working deeper into Physical Design is that many implementation decisions that initially look simple are actually <strong>trade-offs</strong>.</p><p>Clock Tree Synthesis is a good example.</p><p>The question is not simply:</p><blockquote><strong><em>Buffer or inverter?</em></strong></blockquote><p>The better question is:</p><blockquote><strong><em>Which available clock cell gives the best electrical, timing, power, and physical characteristics at this particular point in the clock tree while maintaining the required clock behavior?</em></strong></blockquote><p>That is why understanding CTS beyond just running the tool is valuable.</p><p>The more we understand what the optimization engine is trying to achieve, the easier it becomes to debug <strong>skew, latency, transition, timing, power, and clock-tree quality</strong> when the final results don’t look the way we expect.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db163831b12b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Debugging Setup Timing Degradation from Placement to Post-CTS: A Practical Physical Design Case…]]></title>
            <link>https://medium.com/@sktbttr/timing-degradation-from-place-to-post-cts-physical-design-f5c430b31616?source=rss-50c86b95ce78------2</link>
            <guid isPermaLink="false">https://medium.com/p/f5c430b31616</guid>
            <category><![CDATA[placement]]></category>
            <category><![CDATA[asic-physical-design]]></category>
            <category><![CDATA[soc]]></category>
            <category><![CDATA[timing]]></category>
            <category><![CDATA[vlsi-physical-design]]></category>
            <dc:creator><![CDATA[Sktbttr]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 04:18:06 GMT</pubDate>
            <atom:updated>2026-10-27T10:51:53.337Z</atom:updated>
            <content:encoded><![CDATA[<h3>Debugging Setup Timing Degradation from Placement to Post-CTS: A Practical Physical Design Case Study</h3><p><em>Author: Shaik Thouheed<br>Email: sktbttr@gmail.com</em></p><p>Ever wondered why your timing looks perfectly fine during placement, but suddenly degrades after CTS?</p><p>Let’s understand this through a simple, real-world use case.</p><p>Consider a <strong>1 mm² block (1000 µm × 1000 µm)</strong>. For simplicity, assume a data path enters the block through a port on the <strong>left side</strong>, passes through <strong>two registers</strong>, and finally exits through a port on the <strong>right side</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/405/1*c_XUq6YrdqFIHQIC5j1T3w.png" /></figure><p>Now, let’s add some real numbers to this example. The block runs at <strong>2 GHz</strong>, giving us a <strong>500 ps clock period</strong>, with a single master clock.</p><p><strong>PLACE STAGE:</strong></p><ul><li>IN2REG distance: <strong>200 µm</strong></li><li>REG2REG distance: <strong>600 µm</strong></li><li>REG2OUT distance: <strong>200 µm</strong></li></ul><p>After placement optimization and buffering, the <strong>REG2REG setup path has -30 ps of slack</strong>. At this point, the violation looks marginal and appears recoverable during <strong>CTS or routing</strong>. Based on this assumption, we proceed with CTS.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cmfSXmhmExmL7poM_YgYnw.png" /></figure><p><strong>Post-CTS Stage:</strong></p><p>After CTS, the <strong>REG2REG setup slack degrades to -100 ps</strong>. This looks surprising because the setup slack at the placement stage was only <strong>-30 ps</strong>. Our expectation was that CTS would improve the timing, perhaps bringing the slack closer to <strong>-20 ps or -10 ps</strong>. Instead, it degraded significantly to <strong>-100 ps</strong>.</p><p><strong>So, what went wrong?</strong></p><p>The issue is related to <strong>skew-group management</strong>.</p><p>In general, <strong>I/O registers should be handled separately from the default master-clock skew group</strong>, because their clock-latency/skew requirements can be different from those of the internal registers.</p><p>For example:</p><ul><li>Block skew target: <strong>120 ps</strong></li><li>I/O skew target: <strong>60 ps</strong></li></ul><p>These skew groups need to be properly managed either <strong>by the user or through the implementation flow</strong>.</p><p>Now, consider a case where the <strong>REG2OUT flops are excluded from the default skew group, but the IN2REG flops are not</strong>. In this situation, CTS still treats the IN2REG flops as part of the default skew group and tries to meet the <strong>120 ps block-level skew target</strong> for them.</p><p>This can alter the clock arrival times at the IN2REG launch flops and unintentionally impact the downstream <strong>REG2REG setup paths</strong>, causing the post-CTS timing to degrade significantly.</p><p><strong>But that’s not it.</strong></p><p>We found the skew-group issue and fixed it by excluding the <strong>IN2REG flops from the default master-clock skew group</strong>.</p><p>Now the clock tree looks much healthier:</p><ul><li>REG1 latency: 5<strong>0 ps</strong></li><li>REG2 latency: 6<strong>0 ps</strong></li><li>REG1 → REG2 positive skew: <strong>10 ps</strong></li><li>I/O skew spread: <strong>within 60 ps</strong></li></ul><p>Great. The skew-group problem is fixed.</p><p>But there is still one problem.</p><p>Even with <strong>10 ps of useful positive skew</strong>, the REG2REG path still has <strong>-20 ps of setup slack</strong> post-CTS.</p><p>And here is the real challenge:</p><blockquote><strong><em>The data path has no further optimization scope, and clock push/pull is also not an option.</em></strong></blockquote><p>So now we are stuck with a <strong>-20 ps setup violation</strong>.</p><p><strong>What do we do next?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tJ08McJymMGTLi5C0g3quw.png" /></figure><p>There is one thing, we can bring the REG1 and REG2 closer 600 microns to 500 microns. In this way, our data path will be improved and we can see positive setup slack.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GkqmRhh_Spb5FVVtQ-CKuA.png" /></figure><p><strong>However, there is one caveat:</strong></p><p>With the updated register placement, the <strong>IN2REG and REG2OUT flops are now separated by approximately 250 µm</strong>.</p><p>In general, this is still not the ideal placement. The <strong>IN2REG and REG2OUT flops should ideally be within 100 µm of each other under worst-case conditions</strong>, and in the best case, within <strong>60 µm</strong>.</p><p>Previously, at the placement stage, when the IN2REG and REG2OUT flops were separated by approximately <strong>200 µm</strong>, the I/O timing was:</p><ul><li><strong>TNS: ~7 ns</strong></li><li><strong>WNS: -60 ps</strong></li></ul><p>After the register movement, the distance increased slightly from <strong>200 µm to 250 µm</strong>. However, the timing degradation was significant for IO’s(IN2REG, REG2OUT):</p><ul><li><strong>TNS: -7 ns → -30 ns</strong></li><li><strong>WNS: -60 ps → -150 ps</strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r_0hdpgs3OyVkSkRTu4RiA.png" /></figure><p>There is still some scope to fix these paths during the <strong>ECO stages</strong>. But what if I tell you that this is not a single-port path?</p><p>It is actually a <strong>2054-bit-wide bus</strong>.</p><p>Trying to fix thousands of these paths individually through ECOs is neither practical nor desirable. Such extensive ECO optimization can introduce additional issues, including congestion and timing disturbances on other paths.</p><p><strong>So, is there a better way to address this during P&amp;R rather than relying heavily on ECOs?</strong></p><p>Yes. <strong>Introduce an additional pipeline stage.</strong></p><p>Remember, our block width is approximately <strong>1000 µm</strong>. With the existing pipeline architecture, it is difficult to achieve a well-balanced placement across <strong>IN2REG → REG2REG → REG2OUT</strong>.</p><p>Instead of forcing P&amp;R to close timing on an inherently difficult physical structure, we can provide <strong>RTL feedback</strong> and introduce an additional pipeline register for this bus.</p><h3>How does it help?</h3><p>With the additional pipeline stage, we now have <strong>three register stages between the input and output ports</strong>.</p><p>The physical distribution can now look like this:</p><ul><li><strong>IN2REG distance: ~100 µm</strong></li><li><strong>REG1 → REG2 distance: ~400 µm</strong></li><li><strong>REG2 → REG3 distance: ~400 µm</strong></li><li><strong>REG2OUT distance: ~100 µm</strong></li></ul><p>Now the pipeline stages are distributed much more evenly across the <strong>1000 µm-wide block</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5_tAIgo_XJs4RK9hFFfqZg.png" /></figure><p>Instead of asking one REG2REG stage to cover a large physical distance, the logic is distributed across multiple shorter stages.</p><p><strong>The result? A much more balanced physical implementation, with the setup timing issue eliminated.</strong></p><blockquote><strong><em>Sometimes, timing closure is not about adding another buffer or upsizing another cell. The right solution may be to provide RTL feedback and make the design more physically implementable.</em></strong></blockquote><p>The key takeaway from this example is simple: don’t blindly chase timing violations with ECOs. Understand the physical distance, clock behavior, skew groups, and pipeline architecture first. Sometimes, adding another buffer gives you a few picoseconds, but providing the right RTL feedback can eliminate the problem at its root. That is where Physical Design becomes more than just timing closure.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f5c430b31616" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Exploring SambaNova Systems: Rethinking AI Compute chips with Reconfigurable Dataflow…]]></title>
            <link>https://medium.com/@sktbttr/exploring-sambanova-systems-rethinking-ai-compute-chips-with-reconfigurable-dataflow-99d6453a6345?source=rss-50c86b95ce78------2</link>
            <guid isPermaLink="false">https://medium.com/p/99d6453a6345</guid>
            <dc:creator><![CDATA[Sktbttr]]></dc:creator>
            <pubDate>Sun, 25 Oct 2026 02:45:46 GMT</pubDate>
            <atom:updated>2026-10-25T02:45:46.531Z</atom:updated>
            <content:encoded><![CDATA[<h3>Exploring SambaNova Systems: Rethinking AI Compute chips with Reconfigurable Dataflow Architecture(RDA) (5)</h3><p><strong><em>Author: Shaik Thouheed<br>Email : sktbttr@gmail.com</em></strong></p><h3>From GPUs to a Different Way of Running AI</h3><p>As part of my continued exploration of semiconductor companies and the IP architectures powering the AI era, today I’m looking at <a href="https://sambanova.ai/solutions/neoclouds?utm_source=google&amp;utm_medium=cpc&amp;utm_campaign=&amp;utm_term=sambanova%20systems&amp;utm_content=&amp;hsa_acc=9878570407&amp;hsa_cam=23161114681&amp;hsa_grp=200894964677&amp;hsa_ad=827110487437&amp;hsa_src=g&amp;hsa_tgt=kwd-1405260866073&amp;hsa_kw=sambanova%20systems&amp;hsa_mt=e&amp;hsa_net=adwords&amp;hsa_ver=3&amp;gad_source=1&amp;gad_campaignid=23161114681&amp;gbraid=0AAAAABZUQ-HiKluHudFNM_CN6ld6mfLC_&amp;gclid=Cj0KCQjwrvHWBhDKARIsABmp7h2RonZQdsyt2uEFdWOounV1YaI_WEVlNsNk5VirPwhePGZiglS9fWsaAtWoEALw_wcB"><strong>SambaNova Systems</strong></a>.</p><p>Earlier, I explored CPU and RISC-V companies as well as AI-compute companies, SambaNova is different.</p><p>Instead of developing a conventional CPU or relying entirely on the GPU model of computing, SambaNova has built an architecture around what it calls the:</p><blockquote><strong><em>Reconfigurable Dataflow Unit, or RDU.</em></strong></blockquote><p>The interesting question is not simply whether an RDU is faster than a GPU.</p><p>The more fundamental question is:</p><blockquote><strong><em>What if we organize the hardware around the dataflow of an AI model instead of making the AI model conform to a conventional processor architecture?</em></strong></blockquote><p>Let’s understand that idea.</p><h3>1. Who is SambaNova Systems?</h3><p>SambaNova Systems is an AI computing company founded in <strong>2017</strong> by Rodrigo Liang, Stanford professor Kunle Olukotun, and Stanford professor Christopher</p><p>SambaNova Systems is better understood as a vertically integrated AI infrastructure company than as simply an AI-chip startup.</p><p>The important distinction is that its competitive advantage can come from optimizing the entire stack together, rather than optimizing the accelerator chip in isolation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/784/1*lvKy892-2t-ELiidFHgSvg.png" /></figure><h3>2. The Heart of SambaNova: The RDU</h3><p>RDU stands for:<br><strong>Reconfigurable Dataflow Unit<br></strong>This is the key architectural idea behind SambaNova.<br>To understand why it is interesting, let’s first simplify conventional execution.</p><p>The CPU is instruction-centric. The GPU is designed for massive parallel computation. SambaNova’s Reconfigurable Dataflow Unit (RDU) takes a different approach: <strong>organize computation around the movement of data through a graph of operations.</strong></p><p><strong>1. CPU — Instruction-driven</strong></p><p>Fetches and decodes instructions, then dispatches work to execution units. It is designed for flexible, general-purpose computing, complex control flow, and low-latency tasks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/566/0*s3XNVA-Z9qo9D_4l" /><figcaption>CPU</figcaption></figure><p><strong>2. GPU — Parallel-compute-driven</strong></p><p>Executes many operations concurrently, with specialized matrix engines on modern AI GPUs. Performance depends on parallelism, memory bandwidth, and efficient scheduling.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/733/0*YTmHck8LIYE0frRQ" /><figcaption>GPU</figcaption></figure><p><strong>3. SambaNova RDU — Dataflow-driven</strong></p><p>Maps model operations and data dependencies onto a configurable arrangement of compute and memory resources. Its compiler helps determine where operations execute and how data moves between them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*a3kJtbVR571Pabyo" /><figcaption>RDU</figcaption></figure><p>For <strong>SambaNova</strong>, the design challenge is to make its <strong>compute units</strong>, <strong>programmable memory units (PMUs)</strong>, and <strong>switching fabric </strong>work together efficiently.</p><ul><li><strong>Architecture: </strong>Determines how operations, compute resources, and memory are organized.</li><li><strong>Physical design:</strong> Determines how those resources are placed, wired, clocked, and powered on the actual chip.</li><li><strong>Compiler: </strong>Maps the AI graph onto those resources while considering dependencies, placement, and data movement.</li></ul><p>This is where <strong>SambaNova’s vertical integration</strong> becomes relevant: hardware and software can be designed together around the dataflow execution model.</p><p><strong>One caveat: </strong>a dataflow architecture is not automatically faster than a GPU. The outcome depends on the model, compiler quality, supported operations, memory requirements, batch size, latency target, and measured system-level efficiency.</p><h3>SN40L architecture</h3><p>SN40L is especially useful to understand technically because SambaNova has published more architectural detail about it.</p><p>SambaNova’s research describes this explicitly as:<br><strong>distributed on-chip SRAM → on-package HBM → off-package DDR DRAM.</strong> This is one of SambaNova’s most interesting architectural decisions.</p><h3>Why?</h3><p>LLMs contain enormous parameter sets.</p><p>SambaNova tries to attack that through both <strong>dataflow execution and memory architecture</strong>.</p><p>The company says an SN40L DataScale node can address terabytes of memory and support up to <strong>5 trillion parameters</strong>, including hundreds of models</p><h3>Another important IP: scaling RDUs</h3><p>One accelerator isn’t enough for frontier-scale AI.</p><p>So there is a dedicated RDU interconnect:</p><p>The published SN40L architecture describes a <strong>dedicated inter-RDU network</strong> designed for scale-up and scale-out over multiple sockets.</p><p>And SambaNova now sells integrated rack hardware containing compute, networking, management and power infrastructure.</p><p>So their engineering problem isn’t merely:</p><blockquote><em>“Design an accelerator.”</em></blockquote><p>It is:</p><blockquote><strong><em>Chip → memory → interconnect → node → rack → compiler/software → model serving</em></strong></blockquote><p>That is why I’d classify SambaNova as a <strong>full-stack AI compute company</strong>.</p><h3>SN50 is the important next step</h3><p>SambaNova introduced <strong>SN50 in February 2026</strong>, shifting the company’s positioning even more toward inference and agentic AI.</p><p>An interesting architecture strategy also appeared through its Intel relationship.</p><p>In April 2026 the companies announced a heterogeneous inference blueprint:</p><p>In other words, SambaNova isn’t arguing that an RDU must replace every processor in an AI datacenter. It is exploring systems where each type of processor handles the phase it’s best suited for.</p><p>That is an architecturally sensible positioning.</p><h3>Funding and valuation: very strong momentum in 2026</h3><p>In February 2026 SambaNova announced a <strong>$350M+ funding round</strong> alongside SN50.</p><p>Then, on <strong>July 8, 2026</strong>, it announced the first close of another <strong>$1 billion financing at an $11 billion valuation</strong>.</p><p>The Series F was led by General Atlantic, while TechCrunch reported participation/ties involving major investors including Intel.</p><h3>More important: customers are appearing</h3><p>This matters much more than startup valuation.</p><p>In 2026, <strong>JPMorganChase selected SambaNova as an inference-infrastructure partner</strong>, with SN40L and SN50 intended for secure on-premises AI inference.</p><p>ther publicly identified deployments/customers/partners include OVHcloud, SoftBank and TEPCO Systems, while SambaNova’s press page highlights sovereign-AI deployments and services.</p><h3>Looking at SambaNova through your Physical Design lens</h3><p>This is where I’d suggest you pay particular attention.</p><p>Imagine doing PnR/signoff for this type of chip.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uf-dDMWGrGoqYhfN2qHf7A.png" /></figure><p><strong>What I Learned From SambaNova</strong></p><p>Modern AI-chip design is more than building a fast compute core. It’s about designing the entire system.</p><p>From compute architecture and memory to interconnect, packaging, software, and data centers — every layer influences the others.</p><p>As Physical Design engineers, understanding these connections helps us see beyond timing closure and signoff to <strong>the system-level reason the silicon exists.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DA7jNyWubynfeWtpypYbpw.png" /></figure><p>#PhysicalDesign #VLSI #AIChips #Semiconductor #SambaNova</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99d6453a6345" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Exploring Semiconductor IP: Understanding MatX and Its Approach to LLM Acceleration (4)]]></title>
            <link>https://medium.com/@sktbttr/exploring-semiconductor-ip-understanding-matx-and-its-approach-to-llm-acceleration-4-d35a07400158?source=rss-50c86b95ce78------2</link>
            <guid isPermaLink="false">https://medium.com/p/d35a07400158</guid>
            <category><![CDATA[vlsi-physical-design]]></category>
            <category><![CDATA[llm-acceleration]]></category>
            <category><![CDATA[matx]]></category>
            <category><![CDATA[soc]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Sktbttr]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 09:41:52 GMT</pubDate>
            <atom:updated>2026-10-22T09:41:52.535Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>Author: Shaik Thouheed<br>Email : sktbttr@gmail.com</em></strong></p><p>As part of my journey to broaden my understanding beyond Physical Design and explore semiconductor architecture at a system level, I have been studying companies building interesting semiconductor technologies and IP.</p><p>So far, I have explored companies primarily focusing on RISC-V processor architectures.</p><p>Today, I wanted to look at something quite different:</p><h3>MatX</h3><p>Unlike companies focused primarily on general-purpose CPU architectures, <a href="https://matx.com/">MatX</a> is approaching compute from the perspective of <strong>large language models (LLMs)</strong>.</p><p>And that immediately leads to an interesting architectural question:</p><blockquote><strong><em>What would a processor look like if we designed it specifically for large LLMs instead of trying to make it good at every workload?</em></strong></blockquote><p>That is the idea I found most interesting while studying MatX.</p><h3>The Problem: LLMs Have Very Different Compute Requirements</h3><p>Modern LLM workloads involve an enormous amount of matrix computation.<br>This creates several architectural requirements simultaneously.</p><p>We need:</p><ul><li>substantial compute capability</li><li>efficient data movement</li><li>fast access to model weights</li><li>large memory capacity</li><li>support for long contexts</li><li>communication between many accelerators</li></ul><p>The processor therefore cannot be understood only by asking:</p><blockquote><em>How many compute units does it have?</em></blockquote><p><strong>We also need to ask:</strong></p><blockquote><em>Where does the data come from?<br>Where are the weights stored?<br>Where does the KV cache live?<br>How efficiently can compute resources be utilized?<br>And how do multiple processors communicate?</em></blockquote><p>That is where MatX becomes interesting.</p><h3>MatX One</h3><p>MatX’s processor is called <strong>MatX One</strong>.<br>One of the central architectural ideas MatX describes is its:</p><h3>Splittable Systolic Array</h3><p>Before understanding why it is “splittable,” it helps to understand what a systolic array is.</p><p>A simplified representation could look like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*bJ61gJUF3P_AAdxs07l8mA.png" /></figure><p>Here, <strong>MAC</strong> represents Multiply-Accumulate operations.</p><p>These operations are fundamental to matrix multiplication, which is heavily used in neural-network workloads.</p><p>Instead of continually moving every piece of data back and forth between distant compute and memory structures, systolic architectures allow data to move through arrays of processing elements.</p><p>MatX’s interesting addition is the concept of making this structure <strong>splittable</strong>.</p><p>Conceptually, think of a large compute resource being capable of operating with more flexibility:</p><p>The goal is to combine the efficiency associated with large systolic arrays with better utilization across differently sized and shaped matrix operations.</p><p>For an accelerator, this matters because having enormous theoretical compute capability means little if a workload cannot effectively utilize it.</p><p>Please see the below snippet for more understanding.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*BPOMV_P5jHplDGf8EhHbxw.png" /></figure><h3>Compute Is Only Half the Problem</h3><p>One lesson I continue to see when studying AI accelerators is that:</p><blockquote><strong><em>AI architecture is as much a memory problem as it is a compute problem.</em></strong></blockquote><p>The compute engines need enormous quantities of data.</p><p>Therefore, memory placement and data movement become fundamental architectural decisions.</p><p>This brings us to another interesting part of MatX’s approach.</p><h3>SRAM and HBM Have Different Jobs</h3><p>A useful simplified picture of the architecture is:</p><h3>SRAM</h3><p>SRAM is extremely fast and can provide low-latency data access.<br>MatX’s published architecture emphasizes using SRAM for model weights where low latency is particularly valuable.</p><h3>HBM</h3><p>HBM provides much greater memory capacity and enormous bandwidth. MatX highlights HBM for <strong>KV-cache storage</strong>, helping the architecture support long-context workloads.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/892/1*tn6YiO-87IlmFf0kxQYFjA.png" /></figure><p>MatX’s processor is intended not simply for inference but across major LLM workloads including <strong>training, reinforcement learning, prefill, and decode</strong>.</p><p>That makes the architecture a useful example of how processor design increasingly depends on workload characteristics.</p><h3>Mixture-of-Experts Makes Things More Interesting</h3><p>Another area MatX targets is <strong>Mixture-of-Experts (MoE)</strong> models.</p><p>A simplified way of thinking about MoE is:</p><p>Here’s a technical infographic of MatX’s accelerator IP. It illustrates how Mixture-of-Experts (MoE) models create three key hardware challenges: compute utilization, memory access, and interconnect communication.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*94c9F2uf-jMEiOThlYCDLA.png" /></figure><h3>MatX’s Specialization Is Perhaps Its Most Interesting Decision</h3><p>One reason I find MatX interesting is that the company isn’t positioning the architecture as something optimized equally for every AI workload.</p><p>MatX explicitly says it isn’t targeting areas such as:</p><ul><li>small models</li><li>convolutions</li><li>recommenders</li></ul><p>Instead, its focus is <strong>large LLM workloads</strong>.</p><p>This represents an interesting semiconductor trend.</p><p>Historically, we often wanted processors to become increasingly general-purpose.</p><p>AI is pushing the industry strongly toward <strong>domain-specific architectures</strong>.</p><h3>Looking at MatX From a Physical Design Perspective</h3><p>Being from a Physical Design background, this is where things become especially interesting for me.</p><p>Once we understand the high-level architecture, different questions start appearing.</p><p>For example:</p><h3>Compute density</h3><p>Large arrays of arithmetic units imply highly compute-dense regions.</p><h3>SRAM placement</h3><p>If compute engines need rapid access to data, the relationship between memory and compute becomes an important architectural and physical-design consideration.</p><h3>Power delivery</h3><p>Large amounts of concurrent arithmetic activity naturally make power delivery an important design consideration.</p><h3>Clock distribution</h3><p>Large compute fabrics introduce interesting clock-distribution challenges.</p><h3>HBM connectivity</h3><p>The physical relationship between the accelerator, HBM interfaces, package and overall memory subsystem becomes important.</p><h3>Interconnect</h3><p>High-speed processor-to-processor communication brings another set of PHY, package and system-level considerations.</p><p>I want to emphasize that these are <strong>architectural questions I ask from a Physical Design perspective</strong>, not a claim about MatX’s unpublished physical implementation.</p><h3>Final Thoughts</h3><p>My biggest takeaway from studying MatX is not simply that they are building another AI accelerator.</p><p>It is the importance of <strong>workload-driven architecture</strong>.</p><p>Start with the workload:</p><p><strong>Large Language Models</strong></p><p>Then determine what that workload needs:</p><p><strong>Matrix compute + low-latency memory + high-capacity memory + efficient data movement + interconnect</strong></p><p>And then design the processor and system around those requirements.</p><p>For someone coming from Physical Design, studying architectures like this helps connect the implementation problems we solve every day with the architectural decisions that created those problems in the first place.</p><p>That connection between <strong>workload → architecture → microarchitecture → physical implementation → system</strong> is something I want to continue exploring.</p><h3>Reference</h3><p>For those interested in exploring the company directly: <a href="https://matx.com/">MatX official website</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d35a07400158" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond Physical Design: Understanding Akeana and Its RISC-V Compute Platform (3)]]></title>
            <link>https://medium.com/@sktbttr/beyond-physical-design-understanding-akeana-and-its-risc-v-compute-platform-3-f76dc36727e6?source=rss-50c86b95ce78------2</link>
            <guid isPermaLink="false">https://medium.com/p/f76dc36727e6</guid>
            <category><![CDATA[physical-design]]></category>
            <category><![CDATA[akeana]]></category>
            <category><![CDATA[soc]]></category>
            <category><![CDATA[risc-v]]></category>
            <category><![CDATA[vlsi-design]]></category>
            <dc:creator><![CDATA[Sktbttr]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 03:55:23 GMT</pubDate>
            <atom:updated>2026-10-20T03:55:23.061Z</atom:updated>
            <content:encoded><![CDATA[<p>Author: Shaik Thouheed<br>Email : sktbttr@gmail.com</p><p>As part of my journey toward understanding complete SoC architecture, I have been exploring semiconductor IP companies and trying to understand not only <strong>what IP they provide, but where that IP fits inside a real chip</strong>.</p><p>In this article, I am exploring <a href="https://www.akeana.com/"><strong>Akeana</strong></a>, one of the companies closely associated with the RISC-V ecosystem, and trying to understand its processor IP from a system and SoC architecture perspective.</p><p>After looking at companies such as SiFive and Ventana, today I wanted to explore <strong>Akeana</strong>, another interesting company in the rapidly evolving RISC-V ecosystem.</p><p>The interesting thing about Akeana is that its story doesn’t stop with a RISC-V CPU core.</p><p>Its portfolio spans processors, system IP, coherent and non-coherent interconnect technology, and compute technologies aimed at enabling larger heterogeneous SoCs</p><h3>Starting from the CPU</h3><p>Akeana’s processor portfolio covers a broad performance spectrum.</p><p>Its publicly announced families include the <strong>100 Series</strong>, aimed at microcontroller and real-time applications, the <strong>1000 Series</strong>, with 64-bit cores and MMU support, and the high-performance <strong>5000 Series</strong>, targeting areas including laptops, datacenters and cloud infrastructure.</p><p>That immediately illustrates something important about RISC-V.</p><p>RISC-V isn’t limited to small embedded processors.</p><p>The same ISA ecosystem can extend from MCU-class designs toward high-performance computing systems.</p><h3>A processor alone doesn’t make an SoC</h3><p>This is where my understanding changed while studying companies such as Akeana.</p><p>When we talk about processors, it is easy to focus entirely on the CPU core.</p><p>But a modern SoC may contain many CPU cores together with vector engines, AI accelerators, memory controllers and other specialized processing blocks.</p><p>Those components need an architecture that allows them to communicate efficiently.</p><p>For multi-core CPU systems, cache coherency becomes especially important.</p><p>Akeana therefore complements its processor cores with system and interconnect IP, including a coherent fabric implementing AMBA CHI.</p><p>From an SoC architect’s point of view, the picture therefore becomes much broader:</p><p><strong>CPU → Cache → Cluster → Coherent Interconnect → Memory → Complete SoC</strong></p><h3>Moving toward heterogeneous compute</h3><p>Modern workloads cannot always be efficiently handled by one type of processing engine.</p><p>A system may combine:</p><p><strong>Scalar CPU + Vector processing + Matrix processing + Specialized accelerators</strong></p><p>Akeana describes its platform as supporting scalar, vector and matrix computation while integrating with other accelerator engines.</p><p>Its portfolio has also included an AI Matrix computation engine targeting matrix-multiply offload for AI acceleration.</p><p>This is where RISC-V becomes particularly interesting to me.</p><p>Instead of thinking about CPU architecture in isolation, we can start thinking about <strong>heterogeneous SoC architecture built around an extensible ISA</strong>.</p><h3>Alpine makes the architecture more tangible</h3><p>A useful example is Akeana’s Alpine test chip.</p><p>Akeana says Alpine was taped out in December 2025 as an RVA23-compatible server-class SoC and software-development platform. Its published description includes eight 64-bit out-of-order cores connected through a CHI-based coherent mesh, along with private L1/L2 caches, shared L3 cache and vector engines.</p><p>For me, this is important because it turns discussions about processor IP into something much more concrete.</p><p>It shows how:</p><p><strong>Core IP + Cache hierarchy + Coherent fabric + Vector compute + System IP</strong></p><p>can come together into an integrated chip.</p><h3>Why this matters to a Physical Design Engineer</h3><p>Coming from a Physical Design background, studying architectures like these changes the questions I ask.</p><p>Instead of looking only at:</p><p><strong>floorplan → placement → CTS → routing → timing → signoff</strong></p><p>I am increasingly interested in understanding what architectural decisions created that physical problem in the first place.</p><p>For example:</p><ul><li>Why were these CPU cores grouped together?</li><li>Why does this block need cache coherency?</li><li>What bandwidth does the fabric need?</li><li>Where should memory reside?</li><li>Why is an accelerator separate from the CPU?</li><li>How does architecture influence power, area, timing and congestion?</li><li>How does an architectural requirement eventually become RTL and then a physical implementation?</li></ul><p>That connection between <strong>architecture and physical implementation</strong> is something I want to continue exploring.</p><h3>My biggest takeaway from studying Akeana</h3><p>Akeana helped reinforce one important idea for me:</p><blockquote><strong><em>Modern processor IP is increasingly about more than the processor core itself.</em></strong></blockquote><p>CPU architecture, vector processing, AI acceleration, cache hierarchy, coherent interconnect and system architecture increasingly need to be considered together.</p><p>And RISC-V provides an interesting foundation because of its open and extensible ISA.</p><p>As I continue moving beyond Physical Design toward broader <strong>SoC and Solution Architecture</strong>, these are exactly the kinds of connections I want to understand better.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f76dc36727e6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Physical Design to System Architecture: Understanding Ventana’s Veyron RISC-V CPU (2)]]></title>
            <link>https://medium.com/@sktbttr/from-physical-design-to-system-architecture-understanding-ventanas-veyron-risc-v-cpu-2-be8fda5ab8f5?source=rss-50c86b95ce78------2</link>
            <guid isPermaLink="false">https://medium.com/p/be8fda5ab8f5</guid>
            <category><![CDATA[physical-design]]></category>
            <category><![CDATA[ventanas]]></category>
            <category><![CDATA[vlsi-design]]></category>
            <category><![CDATA[vlsi-solution-arch]]></category>
            <category><![CDATA[risc-v]]></category>
            <dc:creator><![CDATA[Sktbttr]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 13:29:16 GMT</pubDate>
            <atom:updated>2026-10-19T13:47:53.089Z</atom:updated>
            <content:encoded><![CDATA[<h3>Beyond Physical Design: Understanding Ventana’s Veyron RISC-V CPU (2)</h3><p>Author: Shaik Thouheed<br>sktbttr@gmail.com</p><p>Learning how RISC-V CPU IP, multicore compute, system IP and die-to-die connectivity come together in modern semiconductor systems.</p><p>In this article, I am exploring <strong>Ventana Micro Systems and its Veyron RISC-V processor technology</strong>, particularly what makes Veyron interesting from a high-performance computing and system-architecture perspective.</p><p>As part of my journey to expand my knowledge beyond physical design and understand full-chip and system-level architecture, I have been studying different semiconductor IP companies and the problems their technologies are designed to solve.</p><p>One architecture that caught my attention is <strong>Veyron from Ventana Micro Systems</strong>.</p><p>At first, I thought of Ventana primarily as another company building RISC-V processors. But once I started looking at Veyron from a system perspective, the architecture became much more interesting.</p><p><strong>First, RISC-V is not a CPU<br></strong>This was an important distinction for me.</p><p>RISC-V defines an instruction-set architecture. Companies can develop different processor microarchitectures that implement that ISA.</p><p>Ventana’s implementation is the <strong>Veyron CPU family</strong>, designed for high-performance computing. Veyron V2 is positioned for high-performance applications and is available as licensable IP for custom SoCs as well as through a silicon-platform approach.</p><h3>From a CPU Core to a Compute Platform</h3><p>Understanding the CPU core alone isn’t enough.</p><p>A modern compute system may contain multiple CPU cores, several levels of cache, memory-management functions, interrupt infrastructure, accelerators and interfaces to the rest of the system.</p><p>Ventana’s Veyron V2 architecture is designed around high-performance out-of-order CPU cores and a multicore cluster architecture. Its current IP description includes scalable multicore clusters, private L2 caches and shared L3 cache.</p><p>This helped me see the CPU not as an isolated block, but as one part of a much larger compute architecture.</p><h3>Where Chiplets Become Interesting</h3><p>Another interesting aspect of Veyron is Ventana’s emphasis on chiplets.</p><p>Instead of assuming that every function must reside on one large die, a system can be composed from multiple dies where compute and other functionality are separated.</p><p>Veyron V2 was introduced in both IP and chiplet form.</p><p>This approach becomes particularly interesting when thinking about large data-center and heterogeneous-compute systems.</p><h3>Connecting the Compute Chiplet</h3><p>Once a system contains multiple dies, those dies need a way to communicate.</p><p>Ventana’s system-IP portfolio includes <strong>CPX (Cluster Proxy)</strong>, which enables a Ventana compute chiplet to interface with a companion Hub SoC using a D2D transport for AMBA protocols.</p><p>Coming from a physical-design background, I found this particularly interesting.</p><p>We often look at D2D from implementation perspectives such as timing, physical interfaces, routing and signoff.</p><p>But at the architecture level, the more fundamental question becomes:</p><p><strong>Why are we connecting these dies, and what role does each die play in the overall system?</strong></p><p>That shift in perspective is important.</p><h3>CPU IP Is Only Part of the Story</h3><p>A high-performance CPU also needs supporting system infrastructure.</p><p>Ventana’s system-IP portfolio includes technologies for interrupts, memory protection and translation, debug and trace, IOMMU and compute-chiplet integration.</p><p>For example, an IOMMU provides translation and protection for device-originated memory transactions, while APLIC can convert wired interrupts into message-signaled interrupts within the RISC-V interrupt architecture.</p><p>These blocks helped me understand that designing a high-performance processor-based SoC requires much more than selecting a CPU core.</p><h3>My Key Takeaway</h3><p>The biggest learning for me is that we shouldn’t study semiconductor IP in isolation.</p><p>Instead of remembering simply:</p><p><strong>“Ventana makes RISC-V CPUs,”</strong></p><p>I now think about the larger chain:</p><p><strong>RISC-V ISA → Veyron CPU → Multicore Compute → System IP → Chiplets → D2D → Complete System</strong></p><p>That system-level view is something I want to continue developing as I explore opportunities beyond implementation-focused physical design and toward broader <strong>SoC architecture and semiconductor solution-architecture roles</strong>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=be8fda5ab8f5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond Physical Design: My Journey Toward Becoming a VLSI Solution Architect]]></title>
            <link>https://medium.com/@sktbttr/beyond-physical-design-my-journey-toward-becoming-a-vlsi-solution-architect-5ff0d953d2dd?source=rss-50c86b95ce78------2</link>
            <guid isPermaLink="false">https://medium.com/p/5ff0d953d2dd</guid>
            <category><![CDATA[risc-v]]></category>
            <category><![CDATA[vlsi-design]]></category>
            <category><![CDATA[fullchip]]></category>
            <category><![CDATA[vlsi-physical-design]]></category>
            <category><![CDATA[asic-physical-design]]></category>
            <dc:creator><![CDATA[Sktbttr]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 07:16:02 GMT</pubDate>
            <atom:updated>2026-10-19T13:46:24.123Z</atom:updated>
            <content:encoded><![CDATA[<h3>Beyond Physical Design:Understanding SiFive RISC-V IP CPU (1)</h3><p>Author: Shaik Thouheed<br>Email : sktbttr@gmail.com</p><h3>Exploring RISC-V CPU IP, SoC integration, and the bigger picture behind building a chip</h3><p>For most of my career, my view of a chip has been heavily influenced by <strong>Physical Design and Signoff</strong>.</p><p>In this article, I am exploring <strong>SiFive</strong>, one of the companies closely associated with the RISC-V ecosystem, and trying to understand its processor IP from a system and SoC architecture perspective.</p><p>When you work in backend VLSI, you naturally spend a lot of time thinking about questions such as:</p><ul><li>Can this block meet timing?</li><li>How should the floorplan look?</li><li>Can the clock architecture meet skew requirements?</li><li>Where are the congestion hotspots?</li><li>Are setup and hold clean?</li><li>Are we meeting EM/IR requirements?</li><li>Is physical verification clean?</li><li>Can we successfully take this design to tapeout?</li></ul><p>These are important questions.</p><p>But recently I have been deliberately trying to move one level higher.</p><p>Instead of only asking:</p><blockquote><strong><em>How do we implement this block?</em></strong></blockquote><p>I want to understand:</p><blockquote><strong><em>Why does this block exist in the first place, what IP does a complete SoC require, and how should all these pieces be architected together?</em></strong></blockquote><p>That shift has led me to start studying processor architectures, RISC-V, Arm, GPUs, NPUs, NoCs, chiplets, memory systems, D2D interfaces, security IP, high-speed interfaces, and the growing semiconductor IP ecosystem.</p><p>One company I started exploring as part of this journey is <strong>SiFive</strong>.</p><h3>First, RISC-V and SiFive are not the same thing</h3><p>This distinction is fundamental.</p><p><strong>RISC-V is an Instruction Set Architecture (ISA).</strong></p><p>It provides the software-visible instruction architecture on which processors can be built.</p><p>The actual processor implementation involves considerably more:</p><p><strong>Applications / Software</strong></p><p>↓</p><p><strong>Compiler / OS</strong></p><p>↓</p><p><strong>RISC-V ISA</strong></p><p>↓</p><p><strong>CPU Microarchitecture</strong></p><p>↓</p><p><strong>RTL Implementation</strong></p><p>↓</p><p><strong>Synthesis</strong></p><p>↓</p><p><strong>Physical Design</strong></p><p>↓</p><p><strong>Signoff</strong></p><p>↓</p><p><strong>Silicon</strong></p><p>The ISA alone therefore doesn’t define the complete CPU implementation.</p><p>Different processor designers can implement the RISC-V ISA using very different microarchitectures.</p><p>This is where companies such as SiFive become interesting.</p><p>SiFive develops commercial RISC-V processor IP that semiconductor companies can use when building their own SoCs.</p><h3>Why would a company license CPU IP?</h3><p>Suppose I want to build an edge-AI SoC.</p><p>At a very simplified level, my chip may need:</p><ul><li>CPU</li><li>GPU and/or NPU</li><li>Cache subsystem</li><li>SRAM</li><li>DRAM controller</li><li>NoC/interconnect</li><li>DMA</li><li>PCIe</li><li>Ethernet</li><li>USB</li><li>Security</li><li>Debug/JTAG</li><li>Clock and reset subsystem</li><li>Power-management infrastructure</li><li>DFT</li><li>Analog and PHY IP</li></ul><p>Among all of these components, I also have to decide what to do about the CPU.</p><p>I could potentially develop a processor internally.</p><p>But developing a competitive CPU involves much more than simply implementing instructions.</p><p>The processor may require:</p><ul><li>Instruction fetch</li><li>Decode</li><li>Branch prediction</li><li>Execution pipelines</li><li>ALUs</li><li>Floating-point units</li><li>Load/store units</li><li>MMU</li><li>TLB</li><li>L1/L2 caches</li><li>Interrupt handling</li><li>Debug infrastructure</li><li>Multicore support</li><li>Cache coherency</li><li>Verification</li><li>Compiler/software enablement</li></ul><p>Instead of developing everything from scratch, a semiconductor company may choose to license processor IP and concentrate its differentiation elsewhere.</p><p>This is one of the areas where SiFive participates.</p><h3>Where does SiFive fit into the chip?</h3><p>A useful way I have started thinking about CPU IP is:</p><blockquote><strong><em>The CPU isn’t the chip. It is one major subsystem inside the chip.</em></strong></blockquote><p>Consider a hypothetical SoC.</p><p>The CPU subsystem may communicate through the SoC interconnect with:</p><ul><li>NPU</li><li>GPU</li><li>DSP</li><li>system cache</li><li>memory controllers</li><li>PCIe controllers</li><li>Ethernet</li><li>USB</li><li>security engines</li><li>accelerators</li><li>DMA engines</li></ul><p>The overall architecture can be thought of conceptually as:</p><p><strong>CPU cluster</strong></p><p>↓</p><p><strong>L1/L2 cache</strong></p><p>↓</p><p><strong>Coherent NoC / System Interconnect</strong></p><p>↙︎ ↓ ↘︎</p><p><strong>NPU | GPU | Custom Accelerator</strong></p><p>↓</p><p><strong>System Cache</strong></p><p>↓</p><p><strong>Memory Controller</strong></p><p>↓</p><p><strong>DDR / LPDDR / HBM</strong></p><p>And around all of this we still need the supporting infrastructure:</p><p><strong>PCIe + Ethernet + USB + Security + DFT + Debug + Clock + Reset + Power Management</strong></p><p>This is the perspective I find particularly interesting.</p><p>As a backend engineer, I may eventually see an RTL/netlist/hard macro that I need to implement.</p><p>As an architect, however, the questions begin much earlier.</p><h3>The questions become different</h3><p>Imagine a customer says:</p><blockquote><em>“We need to build an edge-AI SoC for industrial vision.”</em></blockquote><p>If I look at this purely from a Physical Design perspective, I may immediately start thinking about:</p><ul><li>process technology</li><li>die size</li><li>macro placement</li><li>utilization</li><li>power grid</li><li>clocking</li><li>timing</li><li>congestion</li></ul><p>But the architecture discussion has to happen before that.</p><p>I first need to ask:</p><h3>Compute requirements</h3><p>What workloads are running?</p><p>Do we need a general-purpose CPU?</p><p>How many cores?</p><p>Do we need a GPU?</p><p>Do we really need a GPU, or would an NPU be sufficient?</p><p>Do we need a DSP?</p><p>Would vector processing help?</p><p>Do we need custom hardware accelerators?</p><h3>CPU architecture</h3><p>What ISA makes sense?</p><ul><li>RISC-V?</li><li>Arm?</li><li>Something else?</li></ul><p>Should the processor be:</p><ul><li>embedded-class?</li><li>real-time?</li><li>application-class?</li><li>high-performance?</li></ul><p>Do we need:</p><ul><li>virtualization?</li><li>vector processing?</li><li>functional safety?</li><li>secure execution?</li></ul><h3>Memory architecture</h3><p>This can completely change the SoC.</p><p>We need to think about:</p><ul><li>L1 cache</li><li>L2 cache</li><li>possibly system-level cache</li><li>on-chip SRAM</li><li>DDR/LPDDR</li><li>HBM for bandwidth-intensive applications</li><li>memory controllers</li><li>PHYs</li><li>ECC</li><li>memory bandwidth</li></ul><p>A powerful accelerator isn’t very useful if we cannot feed it data quickly enough.</p><p>That is one of the key lessons I am taking away from studying modern accelerator architectures.</p><h3>Then comes the NoC</h3><p>Once we have several processing engines, another major question appears:</p><blockquote><strong><em>How does all this data move around the chip?</em></strong></blockquote><p>This introduces another set of architectural topics:</p><ul><li>AXI</li><li>ACE/CHI</li><li>NoC</li><li>cache coherency</li><li>QoS</li><li>bandwidth</li><li>latency</li><li>arbitration</li><li>CDC</li><li>RDC</li></ul><p>And once we move toward multi-die designs, the conversation expands further:</p><ul><li>UCIe</li><li>die-to-die interfaces</li><li>chiplets</li><li>CXL</li><li>PCIe</li><li>SerDes</li><li>advanced packaging</li><li>2.5D/3D integration</li></ul><p>This is an area I find particularly interesting because there is a direct bridge between <strong>architecture and physical implementation</strong>.</p><p>A bandwidth number drawn on an architecture diagram eventually becomes thousands of signals, macros, clock domains, power domains, timing paths, bumps and physical routing resources that the backend implementation team has to close.</p><h3>And finally, everything reaches Physical Design</h3><p>Eventually, the architecture translates into something backend engineers recognize very well:</p><p><strong>Architecture</strong></p><p>↓</p><p><strong>RTL</strong></p><p>↓</p><p><strong>Verification</strong></p><p>↓</p><p><strong>Synthesis</strong></p><p>↓</p><p><strong>DFT</strong></p><p>↓</p><p><strong>Floorplan</strong></p><p>↓</p><p><strong>Placement</strong></p><p>↓</p><p><strong>CTS</strong></p><p>↓</p><p><strong>Routing</strong></p><p>↓</p><p><strong>STA</strong></p><p>↓</p><p><strong>Signal Integrity</strong></p><p>↓</p><p><strong>EM/IR</strong></p><p>↓</p><p><strong>Physical Verification</strong></p><p>↓</p><p><strong>ECO</strong></p><p>↓</p><p><strong>Tapeout</strong></p><p>This is one reason I don’t see my Physical Design background as something I need to move away from to become an architect.</p><p>I see it as the opposite.</p><p>It gives me the ability to connect an architectural decision with its physical consequences.</p><p>For example:</p><blockquote><em>A wider interface might improve theoretical bandwidth.</em></blockquote><p>But what does it do to:</p><ul><li>routing density?</li><li>pin access?</li><li>congestion?</li><li>timing?</li><li>power?</li><li>clocking?</li><li>die area?</li><li>package complexity?</li></ul><p>That’s where implementation experience becomes valuable at the architecture level.</p><h3>Why SiFive interests me</h3><p>Studying SiFive is useful because it forces me to understand several layers simultaneously.</p><p>At one layer there is <strong>RISC-V</strong>, the ISA.</p><p>Then there is the <strong>CPU microarchitecture</strong>.</p><p>Then:</p><ul><li>CPU subsystem</li><li>cache hierarchy</li><li>multicore architecture</li><li>vector processing</li><li>memory</li><li>SoC interconnect</li><li>accelerator integration</li><li>software ecosystem</li></ul><p>And eventually all of that becomes a physical chip.</p><p>That full chain is what interests me.</p><h3>But I don’t want to study only RISC-V</h3><p>This is another important part of my learning journey.</p><p>My objective isn’t:</p><blockquote><em>“Become a RISC-V expert.”</em></blockquote><p>RISC-V is one part of a much larger semiconductor architecture landscape.</p><p>I want to understand how and why architects select among different computing approaches.</p><p>That means studying:</p><h3>CPU architectures</h3><ul><li>RISC-V</li><li>Arm</li><li>x86</li><li>custom processor architectures</li></ul><h3>Parallel computing</h3><ul><li>GPUs</li><li>vector processors</li><li>DSPs</li></ul><h3>AI computing</h3><ul><li>NPUs</li><li>tensor processors</li><li>dataflow processors</li><li>compute-in-memory</li><li>custom AI ASICs</li></ul><h3>Data movement</h3><ul><li>NoC</li><li>PCIe</li><li>CXL</li><li>UCIe</li><li>D2D</li><li>Ethernet</li><li>SerDes</li></ul><h3>Memory</h3><ul><li>SRAM</li><li>DDR</li><li>LPDDR</li><li>HBM</li><li>CXL memory</li><li>emerging memory technologies</li></ul><h3>Supporting IP</h3><ul><li>security</li><li>PLL</li><li>memory controller/PHY</li><li>PCIe/CXL PHY</li><li>DFT</li><li>JTAG</li><li>clock/reset</li><li>power management</li><li>analog/mixed-signal IP</li></ul><p>And increasingly:</p><ul><li>chiplets</li><li>advanced packaging</li><li>silicon photonics</li><li>optical I/O</li></ul><h3>From Physical Design Engineer to VLSI Solution Architect</h3><p>This exploration is also connected to where I would like to take my career.</p><p>I am interested in exploring opportunities around:</p><blockquote><strong><em>VLSI / SoC Solution Architecture, Semiconductor Technical Architecture, and customer-facing technical or pre-sales architecture roles.</em></strong></blockquote><p>I don’t see this as moving away from engineering into pure sales.</p><p>What interests me is the intersection of:</p><p><strong>Deep technical understanding + Full-chip architecture + Customer requirements + Engineering execution + Business solutions</strong></p><p>I would like to eventually be in conversations where a customer says:</p><blockquote><em>“We want to build this semiconductor product.”</em></blockquote><p>And I can help break that requirement down into:</p><ul><li>system architecture</li><li>CPU/GPU/NPU requirements</li><li>memory architecture</li><li>NoC/interconnect</li><li>interfaces</li><li>reusable IP</li><li>third-party IP</li><li>RTL requirements</li><li>verification strategy</li><li>DFT</li><li>physical implementation</li><li>signoff</li><li>packaging</li><li>engineering resources</li><li>project risks</li><li>schedule</li><li>deliverables</li></ul><p>In other words, I want to develop from thinking primarily:</p><blockquote><strong><em>“How do I implement this block?”</em></strong></blockquote><p>toward thinking:</p><blockquote><strong><em>“What chip should we build, what IP and engineering capabilities are required to build it, how should the pieces fit together, and how can we turn the customer’s requirement into an executable semiconductor solution?”</em></strong></blockquote><p>For me, that’s a very exciting direction.</p><h3>My plan from here</h3><p>I’m approaching this as a structured learning exercise.<br>Instead of studying hundreds of random semiconductor companies, I am selecting representative companies across:</p><ul><li>CPU IP</li><li>RISC-V</li><li>Arm-based SoCs</li><li>GPU</li><li>NPU</li><li>AI accelerators</li><li>NoC</li><li>D2D</li><li>chiplets</li><li>CXL</li><li>memory</li><li>security</li><li>SerDes</li><li>photonics</li></ul><p>For each company, I’m trying to answer:</p><ol><li>What problem are they trying to solve?</li><li>What IP or silicon are they building?</li><li>What is the underlying architecture?</li><li>Where does it fit inside a full SoC?</li><li>What does it connect to?</li><li>What memory architecture does it require?</li><li>What software ecosystem is required?</li><li>What does integration look like?</li><li>What challenges does it create for verification and DFT?</li><li>What challenges does it create for Physical Design and signoff?</li><li>What other third-party IP would be required around it?</li><li>What would it take to turn all of this into a complete customer solution?</li></ol><p>SiFive is one of the first companies I am studying through this lens.</p><p>There are many more I want to explore.</p><p>Arm, Tenstorrent, EnCharge AI, Groq, NUVACORE, Tsavorite Scalable Intelligence, Baya Systems, Eliyan, Ayar Labs, and others each represent a different part of the modern semiconductor puzzle.</p><p>And that puzzle is precisely what I want to understand.</p><h3>Closing thought</h3><p>One thing that has become clearer to me is that becoming a full-chip architect doesn’t mean knowing everything more deeply than every specialist.</p><p>That would be impossible.</p><p>A Physical Design architect will always know some implementation topics better.</p><p>A verification architect will know verification better.</p><p>A CPU architect will understand processor microarchitecture at another level.</p><p>An analog architect will understand circuits far beyond my current knowledge.</p><p>What I want to develop is a different capability:</p><blockquote><strong><em>Enough depth in my core area, enough breadth across the rest of the semiconductor stack, and enough system-level understanding to connect all of those specialists into one executable solution.</em></strong></blockquote><p>Physical Design, STA and signoff are where I have built much of my depth.</p><p>Now I am deliberately expanding the breadth.</p><p><strong>The journey is from understanding how to build a block, to understanding how to build the chip, and eventually how to architect the solution around it.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5ff0d953d2dd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding Power State Tables (PST) in UPF/VCLP: A Beginner-Friendly Guide]]></title>
            <link>https://medium.com/@sktbttr/understanding-power-state-tables-pst-in-upf-and-vc-lp-a-beginner-friendly-guide-7fb7a99491eb?source=rss-50c86b95ce78------2</link>
            <guid isPermaLink="false">https://medium.com/p/7fb7a99491eb</guid>
            <category><![CDATA[power-state-table]]></category>
            <category><![CDATA[low-power-design]]></category>
            <category><![CDATA[upf]]></category>
            <category><![CDATA[vlsi]]></category>
            <category><![CDATA[physical-desing]]></category>
            <dc:creator><![CDATA[Sktbttr]]></dc:creator>
            <pubDate>Mon, 07 Sep 2026 16:04:53 GMT</pubDate>
            <atom:updated>2026-09-07T16:23:38.602Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong>Author:</strong> Shaik Thouheed</h4><p>When learning low power verification, one of the first concepts you’ll come across is the <strong>Power State Table (PST)</strong>. Many engineers initially find PST confusing because of the terminology and UPF syntax.</p><p>In this article, I’ll explain PST in the simplest possible way with a practical example.</p><h3>What is a Power State Table?</h3><p>A <strong>Power State Table (PST)</strong> is a table that defines all valid power modes of a design.</p><p>Think of it like a truth table in digital design.</p><p>Just as a truth table defines valid combinations of input signals, a PST defines valid combinations of power supplies.</p><p>The PST helps verification tools such as Synopsys VC LP understand:</p><ul><li>Which power modes are legal</li><li>Which power modes are illegal</li><li>How power domains interact with each other</li><li>When isolation, retention, and level shifters should be active</li></ul><h3>Why Do We Need a PST?</h3><p>Imagine a chip with two power domains:</p><ul><li>CORE Domain</li><li>MEMORY Domain</li></ul><p>Each domain can be either:</p><ul><li>ON</li><li>OFF</li></ul><p>Without a PST, the tool does not know which combinations are allowed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n_1xBeQX1l__Pc_NqE8Psw.png" /></figure><p>The PST explicitly tells the tool which states are legal.</p><h3>PST Creation in Three Simple Steps</h3><h3>Step 1: Define Supply States</h3><p>First, define what ON and OFF mean for each supply.</p><blockquote>add_port_state VDD_CORE -state {ON 1.0} {OFF off}</blockquote><blockquote>add_port_state VDD_MEM -state {ON 1.0} {OFF off}</blockquote><p>Here:</p><ul><li>ON means supply voltage is 1.0V</li><li>OFF means supply is completely shut down</li></ul><h3>Step 2: Create the PST</h3><p>Now create the Power State Table.</p><blockquote>create_pst MY_PST -supplies {VDD_CORE VDD_MEM}</blockquote><p>This command creates a PST containing both supplies.</p><p>You can think of this as creating an empty spreadsheet that will later contain power modes.</p><h3>Step 3: Add Power Modes</h3><p>Now add the valid power configurations.</p><h4>Full Power Mode</h4><blockquote>add_pst_state FULL_ON -pst MY_PST -state {ON ON}</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PyycsJJd7uZLLXkWm_czlw.png" /></figure><h4>Memory Shut Down Mode</h4><blockquote>add_pst_state MEM_OFF -pst MY_PST -state {ON OFF}</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wLBQkrVrVLBVfIxqsg4kJg.png" /></figure><h4>Complete Shutdown Mode</h4><blockquote>add_pst_state SHUTDOWN -pst MY_PST -state {OFF OFF}</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xqu99-D7wygl_hZRlCwo_w.png" /></figure><h3>Visual Representation</h3><p>Our final PST would look like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j0r9PUe1UT64acKSNZ8NMw.png" /></figure><h3>Easy Trick to Remember PST Creation</h3><p>Whenever you create a PST, remember this formula:</p><blockquote><strong>Supply States-&gt;Create PST-&gt;Add PST States</strong></blockquote><p>(OR)</p><blockquote><strong>add_port_state-&gt;create_pst-&gt;add_pst_state</strong></blockquote><p>If you remember these three commands, PST creation becomes very easy.</p><h3>Real-Life Analogy</h3><p>Think about a laptop.</p><p>It can operate in multiple modes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gkdBaFz2a0AEohSKukiYjw.png" /></figure><h3>Final Thoughts</h3><p>A Power State Table is nothing more than a list of valid power configurations in your design. Once you understand that each row represents a legal power mode, PST creation becomes straightforward.</p><p>Remember the golden rule:</p><p><strong>Define supply states → Create PST → Add valid power modes.</strong></p><p>Mastering PSTs is an important step toward understanding more advanced low-power concepts such as Isolation, Retention, Level Shifters, Power Shutoff, and UPF-based verification.</p><p>Happy Low Power Verification!</p><p><strong>Author:</strong></p><p>Shaik Thouheed (sktbttr@gmail.com)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7fb7a99491eb" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>