@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20260601

Episode Date: June 1, 2026

- Huawei's Tau Scaling, LogicFolding: breakthrough or brilliant positioning? - China’s new tech disclosure strategy continues - Wafer-scale chips in China? - Snowflake’s $6B AWS deal and what it says about AI inference demand - The Return of CPU Computing - NVIDIA’s CPU-only Vera-based Supercomputer and the rise of agentic AI infrastructure - OpenAI math breakthrough and what it says about AI-assisted discovery [audio mp3="https://orionx.net/wp-content/uploads/2026/05/HPCNB_20260601.mp3"][/audio] The post HPC News Bytes – 20260601 appeared first on OrionX.net.

Welcome to HPC Newsbytes, a weekly show about important news in the world of supercomputing, AI, quantum computing, and other advanced technologies. Hi, everyone. Welcome to HPC Newsbytes. I'm Doug Black, and with me is Shaheen Khan. Huawei claimed last week at an I-Tripple E event in China, what Tom's hardware described as, quote, a sanctions-busting breakthrough, unquote, of being able to produce 1.4 nanometer class chips by 2031, with 55% higher transistor density. This is Huawei's new logic folding chip architecture that the company says can bypass restrictions

on China's access to extreme ultraviolet chipmaking machines those produced by the Dutch company, ASML. Sheen, I'm very interested in your thoughts on Huawei's claim that their Tao scaling law, T-A-U, will replace Moore's law. Is there validity here, do you think? Well, as usual, it helps to see this through the lens of Huawei, the company, and also China, the country. The big announcement and stage presentation is consistent with what we said a few weeks ago, as China's apparent change in strategy to disclose more about their technology advances.

I would not be surprised if they started participating in the top 500 list again this year. What Huawei has done is excellent marketing. They have identified two valid and well-accepted set of strategies, namely a focus on reducing signal delays via 3D design and packaging, package them into nicely labeled brands, positioned them as industry-level laws, and provided a time frame based not on what lithography can do today, but on next-generation lithography. That is a media-friendly storyline, and it's been duly picked up and amplified. But big picture, things carry on as before, and nothing has changed.

China does not have access to advanced lithography. Advanced lithography is needed. It will not be able to develop it indigenously anytime soon and must therefore make up for it through other parts of system design. And that means holistic co-design and co-optimization across the full stack, which is what everyone else is doing too. Now, efficiency drops as the number of nodes grows and slower chips require a larger number of chips and nodes and racks for a given performance. So Huawei would need to optimize across the stack in a tougher optimization zone.

It's usually a bigger and harder task than what you have to do with clusters based on faster chips, but it's been done before. When you look at the entire software stack and algorithms for learning and inference and apps on top of them, especially with today's learning algorithms that rely on heavy communication and shared memory constructs and contrast that with what the hardware needs to do, advances in those softer areas will be more imminent, and of course that's something that the entire industry is pursuing as well. At the chip level, another path for China could be wafer scale chips. That would reduce the number of nodes required for a given overall performance. The Chinese Academy of Sciences has a development project in that area, and that may very well feed into Huawei, but it is not discussed in this announcement.

There are a lot of problems to be solved for wafer scale chips that we have discussed on this podcast before, like connectivity among different parts of the wafer, heat distribution and structural integrity, yield management and routing around failures, etc., etc. It took Cerberus and TSM years to solve them, so that is a difficult path, too. The AI landscape is changing fast, and with it, the compute needs and sources of compute. News broke last week that Cloud Data Storage Company's Snowflake signed a $6 billion five-year deal with Amazon Web Services for AWS AI inference chips, i.e. Gravaton, arm-based CPUs. The deal in part is due to the large amount of corporate data housed in Snowflake Storage on AWS. It also reflects the shift in the AI market toward inference, the move from training, then to daily usage, and now to automation via AI agents,

which means CPU usage is exploding. The announcement came alongside a blowout Q1 earnings report for Snowflake, which beat revenue estimates and raised its full year guidance, which led the stock to surge by 30, Now, in its fifth generation, Graviton 5 can provide 192 cores and deliver some 25% more performance than Graviton 4. Not surprisingly, AWS's Andy Jassy says their CPUs offer plenty of compute at lower price, and apparently, Snowflake agrees. The announcement is not quite just the subscription renewal spun by the vendors, but it is positioned well and spun well nevertheless.

it indicates massive demand for infrastructure and includes CPUs and GPUs, but with heavy emphasis on CPUs. So it simultaneously indicates growth for the companies, shows further adoption of AWS' homegrown graviton CPUs, and sends a signal that going forward, GPUs are maybe not as critical as they have been. In reality, market demand for all technologies is growing and the return of CPU-based computing is a welcome release valve. As for the size of the deal, for the full year 2025, AWS generated $128.7 billion in total revenue. By the end of Q4 2025,

AWS's growth was 24% year over year, pushing its annualized revenue run rate to $142 billion. So $6 billion over five years is obviously huge and one of the biggest deals of its kind, but numbers are big in all sorts of places these days. Snowflake runs on Azure and Google Cloud too, but AWS is the main provider. It was also reported that Snowflake had generated some $7 billion through the AWS marketplace over its history and customers' own spending through AWS on Snowflake,

presumably ramping up fast over the years, also would likely count. And given that software is sticky and many of those customers would continue, and others would join them, the deal seems to be a calculated projection. The Snowflake has also announced multi-year deals with Anthropic and Open AI on the order of $200 million each, along with collaborations with meta-AI and mistral, the French AI company.

And given AI's rapid growth, those deals could grow as well. As we're seeing, the surge in AI inference has led to more companies producing new chips to compete in the CPU market, and that includes the inventor of GPUs. We're talking, of course, about NVIDIA. CEO Jensen Wong announced earlier this month the intent to grow NVIDIA's CPU presence with a chip called VARA, designed specifically for AI. The new VERA CPU rack integrates 256 liquid-cooled Veras and can sustain more than 22,500 concurrent CPU environments. AI factories can quickly deploy and scale to tens of thousands of simultaneous instances and agentic tools in a single rack, according to the company.

And the company said the chip delivers results with twice the efficiency and 50% faster than traditional CPUs. Well, we've been covering the return of CPU-based computing for a while now. A few weeks ago, we described the growing importance of CPUs for orchestration of agentic AI and their on-chip accelerators and therefore use and inference to cut costs. We also talked about China's CPU-only supercomputer a few weeks ago. And we have described Fujitsu's Fugaku and fugu.net next at length with the upcoming Monaco CPU and ability to use GPUs 2 when needed.

So CPU computing is coming back. And fortunately for Invidia, it's taken it long enough that Nvidia can play. It will be a competitive market that includes Intel and AMD, of course, but also Arm itself with a chip they called AGI, no less, an ampere, which seems to be focusing on large deployments, and established homegrown CPUs by AWS, Microsoft, and Google, and then Risk Five players like Sci-5, Ventana, Esperanto, SoftGo, and Alibaba, which can do Risk Five and Arm,

especially through its chip design subsidiary T-Head, T-Hash, head. We'll close with the Wall Street Journal report last week about an open AI model, cracking a well-known math problem, well-known in math circles, that has resisted a solution for 80 years. The problem looks to me something like an optimization problem. The journal article describes it as, quote, if you put n dots on a sheet of paper, how many pairs of dots can be exactly one unit apart? The open AI model solved it and, leading mathematicians are stunned. No less than Timothy Gowers, professor at College de France, and winner of the prestigious Fields Medal, called it a milestone in mathematics.

This achievement is consistent with what renowned mathematician Professor Terence Tao of UCLA has said about the role of AI and math. Tao sees AI as a powerful tool that transforms mathematics by lowering the cost of exploration and accelerating routine workflows. AI excels at large-scale meta-searches and validating, quote, crazy ideas across infinite combinations. But he cautions that raw computer-generated proofs often miss the journey and miss important conceptual insights. And of course, AI remains prone to mistakes, including very subtle, logical hallucinations. So he sees a critical need for formal, verifiable. tools. All that would transform the mathematician to intuitive director of automated logic.

Shaheen, I'll repeat a point that's often made, which is that AI is permanently emerging into something more powerful. So if models are solving problems of this sort now, where will it be in 10 years in 25? Yeah, absolutely. Absolutely, it's coming. Well, as you all know, I believe AI delivers amazing results when A, it has information that humans gave it, but did not know they were giving it, i.e. the information content of what AI ingests is higher than what humans perceive. Or B, you need very high iteration count to incrementally build towards the result, and computers can iterate billions of times. Or C, the result requires the right combination of several things from among billions of possibilities, and AI can sift through them all.

What Open AI has done here is a notable and sophisticated version of this combinatorial approach. Now, a human would have to do what I call a, quote, perceptive shot in the dark, end quote, because the meta space of combinations is so large, the human brain has to use intuition and do some extreme filtering and focus on a few possibilities. When that works, it's genius. When it doesn't, it's a series of dead ends. This is why a quote, crazy thinker can sometimes find the answer by considering a novel unexpected combination. But what AI has done here is described as more than just a sophisticated search.

Even the most powerful supercomputers are set to be easily paralyzed by what they call the combinatorics of pure math, meaning that calculating every combination of mathematical fields is just not possible. The Open AI model was able to consider that right combination of mathematical concepts to begin with and then applied the right logic to synthesize them. It had to use something similar to learned intuition, smart filtering, assigning probabilities to eligible mathematical concepts, and then evaluate many more combinations than a human could. So AI was able to produce a very good short list, which is the impressive part,

even though that list is way, way longer than a human's short list, the chain of thought it produced was 125 pages long, covering many explorations, each of which could take a long time for a human. To paraphrase Professor Tao, AI can be richer and broader, but not deeper. All right, on that promising note, that's it for this episode. Thank you all for being with us. HPC Newsbytes is a production of Orion X, Shaheen Kahn and Doug Black host the show. Every episode is posted on Orionx.net. If you like the show, please rate and review it.

Thank you for listening.