←
Chad Wallace944 posts
Chad Wallace avatar

Chad Wallace

@siliconcodesign

Deep-Tech Analyst, MSEE @JohnsHopkins. Expert-driven system architecture deep-dives of AI datacenter hardware at siliconcodesign.com/

◷ Joined April 2025▧ 138 Photos and videos
Chad Wallace avatar
Chad Wallace@siliconcodesign·Sep 23

Most people who use AI and invest in AI Infrastructure operate at a high abstraction level and chase trends. However, developing a deep and broad understanding of AI Infrastructure is difficult because domain-specific technical data lives in isolated silos and isn't translated well to a broad audience. As a result, engineers often don't have system context of what they are designing and investors often don't have a decent mental model of what they are investing in. My goal is for investors and engineers to think more like system architects, because architecture decisions have financial implications and influence day-to-day execution of engineering teams. Over the past year, I've travelled to the semiconductor industry's flagship conferences and learned how these subsystems worked from the industry's foremost experts. I've written a complete suite of mental models on the subsystems that comprise AI Infrastructure: - Mixed-Signal ICs (SerDes, ADCs, PLLs, Bandgaps) - High Speed Optical/CPO - High-Power PMICs - High-Speed Signal Integrity - Advanced Packaging - AI for Chip Design / EDA - AI Accelerators - Memory (HBM and SRAM) Link to the website is in my bio and in the comment below 👇

Post attachment
Chad Wallace avatar
Chad Wallace@siliconcodesign·Nov 17

Halfway through @TomorrowXSummit and I want to share a lecture series I found extremely helpful while preparing: "Stanford's MS&E435: Economics of the AI Supercycle" Many of the guest lecturers there were speakers at @TomorrowXSummit I learned of for the first time, such as @alighodsi, @altcap and @ChaseLochmiller. youtube.com/watch?v=LNSvp-9b…

Chad Wallace avatar
Chad Wallace@siliconcodesign·Nov 16

There are a lot of multi billion dollar funding announcements and market uncertainty in the AI infrastructure buildout, but most explanations only cover one aspect of the big picture. What's missing is a holistic understanding of how the financing and technology fits together. Ahead of @TomorrowXSummit, @NeurIPSConf, and @AnthropicAI's expected IPO, I put together a comprehensive overview of the financial status of AI Infrastructure and the risks involved. This post highlights the scale of CapEx spending of several key players with trailing twelve month (TTM) financial numbers and new reports of multi-billion dollar funding/partnership announcements in the last year. Significant players in the AI Infrastructure buildout are represented in this post, including: -AI Model providers (@OpenAI, @AnthropicAI), -Fabless Semiconductor Companies (@nvidia, @Broadcom, @MarvellTech, @AMD) -Memory/Logic Fabs (@SKhynix, @Samsung, @MicronTech ), -Hyperscalers (@Microsoft, @amazon, @Meta, @Google, @Oracle, @SpaceX) -Neoclouds (@CoreWeave, @nebiusai), -Energy providers (@SBEnergy) -Financiers (@apolloglobal, @BlackRock). Whether or not you interpret financing as "circular", this post shows how all of these players are interrelated to each other. This way, you can understand the scale of where the money is coming and going. siliconcodesign.com/p/the-ai…

Chad Wallace avatar
Chad Wallace@siliconcodesign·Nov 15

The mixed signal design process revolves around adding levels of fidelity to architectures to detect problems at the highest possible abstraction level before handing blocks off to designers. A lot of inexperienced mixed-signal designers think that the design process starts with block level design, but this is backwards. Mixed-signal designs begin with behavioral modeling of a high-level system model, whether it be through VerilogAMS, MATLAB/Simulink, or some other high level language. Then, levels of fidelity are added to catch system-level sensitivities - including noise, bandwidth, and mismatch - before specifications are handed off to individual block owners. Unfortunately, most schools only really teach design from the bottom-up with transistor level physics and block-level topologies. This isn't wrong per se, but it often leaves out a lot of high-level modelling that comes with industry experience from working on products in the market. Understanding the overall design process is becoming much more important when competitive, market-defining mixed-signal solutions are becoming more tightly integrated, such as PMICs, audio, and ADCs. Such systems requires block-level designers and verification engineers to understand the system architecture process flow to coordinate and execute efficiently within it. I wrote a post that provides a high-level overview of the mixed signal design flow from system architecture to transistor level design. siliconcodesign.com/p/a-surv…

Chad Wallace avatar
Chad Wallace@siliconcodesign·Nov 15

If you want to know who is bottlenecking the AI compute buildout, stop looking at the AI researchers and start looking at the packaging engineers. Many people take electronic packaging for granted as putting a container around a chip. However, electronic packaging is a sophisticated mechanical and material science problem that is becoming increasingly critical to understand to squeeze more performance out of AI compute. Today, TSMC's Chip-on-Wafer-on-Substrate (CoWoS) is considered the package of choice for HPC compute since it enables dense interconnects. However, in its early days, it was very expensive and customer adoption was very slow at first, leading many people to jokingly refer to it as "NoWoS". Nowadays, there are other integration alternatives such as Intel's EMIB technology and glass substrates. All of these options face first-principles yield challenges such as CTE and mechanical stability. I wrote a comprehensive post that covers the foundations of advanced packaging, from packaging classifications, types of solder bumps, and much more. siliconcodesign.com/p/a-mast…

Chad Wallace avatar
Chad Wallace@siliconcodesign·Nov 14

Mixed signal design is one of the most complex, yet inefficiently taught subjects amongst Electrical Engineer Subspecialties. Mixed signal design is arguably the "deepest" EE specialty that interfaces with a lot of other specialties, such as DSP. However, mixed signal design suffers from a number of issues that often times drives perfectly capable talent away from it: -Long chip feedback time due to the length of time it takes for engineers to get data back and measure against expected simulated performance -The high learning curve for complex simulation software such as Cadence with advanced features in a UI that still looks like it's from the 1980's -The high expense of chip tapeouts, making new hires inherently risky -Professors that often don't have access to leading edge process nodes due to NDAs There are several initiatives helping to lower this barrier to entry and develop talent such as @ieee_isscc's Circuit Insights, Razavi's books, and AI. For those without access to these tools, I wrote a blog post how how to self-teach analog design better to stand out in interviews by constructing a hierarchal "mental model" of key concepts and understanding how they build on top of each other. siliconcodesign.com/p/learni…

Chad Wallace avatar
Chad Wallace@siliconcodesign·Nov 13

You can have the most powerful AI accelerator in the world, but it's meaningless if it can't move data efficiently. The question is, why is data movement so expensive, and how do chip architects design around it? I wrote a first-principles primer that explores these questions, covering MAC operations, SRAM vs DRAM, model & accelerator performance metrics, and "Eyeriss", a classic AI accelerator that ties these concepts together. siliconcodesign.com/p/ai-acc…

Chad Wallace avatar
Chad Wallace@siliconcodesign·Nov 12

AI didn't go from Atari to ChatGPT in a straight line. Behind frontier AI models are competing labs with different philosophies and a paper trail of landmark AI papers that built on top of each other. I examine that messy history and break down key papers that influenced frontier AI models today: siliconcodesign.com/p/a-brie…

Chad Wallace avatar
Chad Wallace@siliconcodesign·Nov 11

As AI compute racks become much more powerful, power delivery, liquid cooling, and high-speed networking are all running into difficult engineering constraints simultaneously. Often times, solving a problem in one subsystem creates new challenges elsewhere. At @OpenComputePrj's Global Summit, I studied how these tradeoffs are evolving across three critical systems: - 800VDC Power Delivery: More energy storage near the compute helps smooth dynamic AI workloads, but also increases fault current and makes protection harder - Liquid Cooling: Finer cold plate microchannels can improve thermal performance, but increase the pressure drop and CDU pumping requirements - High-Speed Networking: As electrical speeds approach 448Gbps+, all aspects of signal integrity and reliability become increasingly difficult to optimize across chips, packages, and cables OCP 2026 democratized a lot of technical information and engineers need to understand the tradeoffs and constraints of adjacent systems, rather than just their own. This deep-dive, along with my entire website, connects the power, cooling and networking architectures behind AI clusters, drawing from insights from several experts at OCP 2026. siliconcodesign.com/p/ocp-20…

Chad Wallace avatar
Chad Wallace@siliconcodesign·Nov 10

Everyone talks about the hundreds of billions being spent on AI infrastructure. But smart investors have to ask: Where does the revenue come from to justify these investments, and what happens if it falls short? Answering this question requires looking beyond headlines and connecting the dots amongst domains often analyzed in isolation: AI demand, GPU economics, semiconductor supply, data center CapEx, and institutional financing. In my latest deep dive, “The AI Infrastructure Buildout: Who Pays, Who Profits, and What Could Go Wrong” I provide a system-level framework to understand the economics and financial risks behind the AI infrastructure build out, including: -The scale of the AI data center buildout -AI Compute economics -The financial position of major players: AI model providers, fabless semiconductor firms, manufacturers, hyperscalers, and neoclouds -Six major investment risks: from hardware obsolescence to execution delays I’ll be at @TomorrowXSummit in Austin, TX on Nov 17-18 where institutional investors and AI infrastructure companies will be discussing many of these challenges. siliconcodesign.com/p/the-ai…