Jun Yang

*** Cross-die links introduce skew, coherence issues, and latency that don’t appear in monolithic SoCs. *** Chapter 5 of my book on chiplets details die-to-die interconnects (UCIe, BoW, AIB, and others) and the hooks needed to validate them. Now available on Amazon in hardcover, paperback, and Kindle. Links in the comments.

19

1 Comment

I work closely with several leading autonomy and robotics companies in the Bay Area 🚗🤖 and regularly connect with engineers across planning, perception, simulation, and robotics systems. One trend I’m seeing across many teams: strong systems engineers are increasingly in demand. The stack (often C++ / Python) matters, but what matters more is deep domain experience in robotics or autonomous systems, particularly around safety, evaluation, validation, and fault handling. If you’re working in this space or curious about what’s happening across AV and robotics, feel free to reach out. #robotics #autonomoussystems #physicalAI

28

4 Comments

Over the past few months I've been trying to get deeper into RTL and understand how real hardware blocks inside processors are built. One thing that caught my attention while reading about modern CPUs and GPUs from companies like AMD and Nvidia is how much they rely on hardware performance monitoring. These chips have internal counters that track different micro-architectural events so engineers can analyze how the processor behaves under real workloads. That idea was really interesting to me, so I decided to try building a simplified version of a Performance Monitoring Unit myself. I implemented a parameterized PMU in SystemVerilog that monitors multiple event signals and tracks them using independent saturating counters. The design is modular, scalable, and verified through simulation with assertions and a testbench. This project was a great learning experience for me because it pushed me to think more about how hardware blocks are structured, not just how to write Verilog that compiles. Still a lot more to learn in RTL and silicon design, but building things like this is honestly the most fun part of studying computer engineering. Repo if anyone wants to check it out: https://lnkd.in/ejRKvtcm

58

3 Comments

The number of ML papers submitted has grown exponentially, overwhelming editors and program committees. This leads to tons of bogus papers getting published (which then infect later papers with the bogus content). Good papers get pushed out. Making a real contribution is hard, and there’s usually only modest gains. But if a paper is not constrained by fact, incredible “improvements” are easy to claim.

13

3 Comments

Dark horse emerges in the AI chip race, Nvidia's 4 trillion market cap red line under threat On September 5, Nvidia's stock closed down 2.7% at $167.02, with its market value remaining above $4 trillion despite an intraday drop of more than 4.4%. This price level was last seen on July 14. Meanwhile, Broadcom's stock surged 9.41% to $334.89, reaching intraday highs of $356.24 and pushing its market capitalization to $1.575 trillion, surpassing Saudi Aramco. Broadcom recently reported third-quarter results that beat expectations and provided strong guidance for the fourth quarter, driven by an accelerating AI chip business. On the earnings call, Broadcom CEO Hock Tan revealed a surprise $10 billion custom chip order from a previously undisclosed customer, believed to be OpenAI. The chip, designed jointly by Broadcom and OpenAI, is expected to debut next year to reduce dependence on Nvidia products. This development positions Broadcom as a more direct competitor to Nvidia in AI hardware. Industry experts, including Clearstead Advisors' Jim Awad, suggest that while Nvidia may face increased competition, the rapid expansion of the AI market could allow it to maintain steady growth despite potential market share losses. Awad noted the recent Nvidia stock decline is likely a knee-jerk reaction and emphasized watching how Broadcom and OpenAI execute their plans. Nvidia’s share price has declined about 10% from its August high, cutting roughly $470 billion in market value, yet it remains the world's most valuable company, with Microsoft trailing at about $3.7 trillion. Mizuho Securities' Daniel O'Regan highlighted Nvidia's relative low compared to Broadcom not seen in 18 months and noted that Broadcom's stock has outperformed Nvidia year-to-date. Investors are broadening their focus to other AI players, with Broadcom emerging as a new attractive contender in the sector. #Semiconductors #China https://lnkd.in/gC6cCT7e

1

From Kernel to User Mode: The Ring 3 🦀Lets Run GTA 5 in JC-OS 😎 #RustLang #OsDev #SystemsProgramming #X86_64 #JCOS #TheRustEagle

2

This paper brings clarity to the AGI , offering a rigorous framework for measuring progress. But it also highlights major gaps in today’s AI systems: Long-Term Memory Storage (MS): Current models (like GPT-4 and GPT-5) cannot reliably store new information over time. They suffer from “amnesia,” forcing them to relearn context in every session. Long-Term Memory Retrieval (MR): AI can retrieve facts quickly, but often hallucinates or confabulates, making retrieval unreliable. Working Memory (WM): Large context windows are used as a workaround, but this is inefficient and doesn’t scale for tasks needing persistent context. Capability Contortions: Practitioners rely on external search tools and context windows to mask these memory deficits, but these are not true solutions.   The authors argue that until AI systems can reliably acquire, store, and recall information like humans, claims of AGI are premature. This paper is a good read for anyone interested in the future of AI and what it will take to reach true general intelligence.   Read more: https://lnkd.in/g4v9bc3u   #AI #AGI #ArtificialIntelligence #TechResearch #Memory #FutureOfAI

8

1 Comment

Physical Intelligence turns terabytes (TBs) of robotics sensor and video data into high-value training datasets. Before Ray on Anyscale, scaling multimodal data pipelines across diverse hardware slowed progress and demanded deep infrastructure expertise, pulling developers away from differentiated research work. With Anyscale, Physical Intelligence enables more than a dozen of its developers to run experiments concurrently and deploy workloads seamlessly across a multi-cloud environment, accelerating iteration and keeping teams focused on their mission to build advanced AI for robotics. Read the full story: https://lnkd.in/g29uj4vi

29

The Future of AI Infrastructure: Open, Heterogeneous, and Scalable Insightful overview by Techgoondu. on why the next major leap in AI won’t just be about faster GPUs — but about heterogeneous and open systems. Instead of locking customers into one proprietary stack, AMD is pushing a more open compute foundation that combines CPUs, GPUs, networking, memory, and software into balanced, scalable platforms tailored for today’s diverse AI demands. Key takeaways: 🔹 AI infrastructure is expanding beyond GPUs — customers need high‑performance CPUs, networking, and storage that work together seamlessly. 🔹 Open standards and software (e.g., ROCm) give organizations flexibility to innovate and scale without vendor lock‑in. 🔹 AMD’s upcoming Helios platform brings EPYC CPUs, Instinct GPUs and Pensando networking together with ROCm to simplify large‑scale AI deployments across cloud, edge and enterprise environments. This is a timely reminder that the evolution of AI infrastructure is as much about architectural openness and system design as it is about raw compute. The future of AI won’t be one‑size‑fits‑all — it will be flexible, interoperable, and ready for tomorrow’s workloads. Full Overview here: https://lnkd.in/gQzU4N-z #AI #AIInfrastructure #AMD #HeterogeneousComputing #AMDBrandAmbassador #OpenEcosystems

110

30 Comments

ExecuTorch 1.0 is here - bringing a unified PyTorch workflow to billions of Arm-based edge devices. For those building at the edge, that means: 🚀 Faster, simpler development and deployment 📱 Greater reach for apps and workloads ⚡ Higher performance and efficiency across Arm CPUs, GPUs, and NPUs With Arm KleidiAI, CMSIS-NN, and TOSA integrations in ExecuTorch, it’s now easier than ever to bring high-performance AI to life everywhere.

9

The most interesting paper I read last week, showing promising work on optical generative models that offload parts of the compute from electrons to photons. UCLA researchers demonstrated generating images by physical light propagation rather than digital computation. The approach maps diffusion model concepts to free-space optics: a shallow digital encoder converts random noise to phase patterns, then a trained diffractive surface processes light to generate images with the optical computation happening in under a nanosecond (overall speed is limited by the SLM refresh rate). This cross-pollination between fields (GenAI → Optics → GenAI) strengthens the entire research ecosystem. Moving from bits to light as the medium opens new possibilities for energy-efficient and significantly faster inference. Paper: https://lnkd.in/gC4By9Vv

39

1 Comment