@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20240826

Episode Date: August 26, 2024

- Hot Chips 2024 conference - LANL electricity needs enter the senate race - A small modular nuclear reactor in a 30-inch borehole a mile under? - AMD in $4.9B deal to buy ZT Systems [audio mp3="https://orionx.net/wp-content/uploads/2024/08/HPCNB_20240826.mp3"][/audio] The post HPC News Bytes – 20240826 appeared first on OrionX.net.

Welcome to HPC News Bites, a weekly show about important news in the world of supercomputing, AI, and other advanced technologies. Hi, everyone. Welcome to HPC News Bites. I'm Doug Black of Inside HPC, and with me is Shaheen Khan of OrionX.net. The annual Hot Chips Conference is being held this week at Stanford, as in the past. It's become a bigger and more prominent conference as interest in chips has grown globally, and also

because it can provide a sneak preview of future chips. I recall last year how UCIE, new AI chips, and new types of processing like NPU, DPU, XPU, and energy efficiency were among the hot topics. They continue with those this year as well. The structure of the sessions are as follows. AI-assisted hardware design, the cooling of hot chips and thermal technology, high-performance processors, specialized processors, which includes some AI-focused systems and memory, but also AI processors show up as a category and covers a half-dozen chips, and finally, networking processors. It's a hybrid conference, so you can register to attend online. They also usually make the presentations available a few months after the conference.

Always good and important conference. Energy requirements remain front and center for existing data centers that need to be modernized and new ones that need power lines. Case in point is Los Alamos National Lab, which has proposed to build a new 115 kilovolt transmission line. That translates to something like 100 like 150 megawatts, if my math is correct. But there's a catch. The line would have to cut across 14 miles of protected wilderness. Right. They need the additional power for their next generation Exascale Class ATS-5 supercomputer scheduled for the 2027 timeframe. Complexities in generating power are well known, but transmitting power is another big challenge. The matter has now entered state politics and become a topic

in New Mexico's U.S. Senate election, with the incumbent appearing to push for alternatives and the challenger saying time is of the essence and the project should be green-lighted. The bigger topic here, of course, is the need for cheaper, carbon-friendly sources of energy, along with attracting novel approaches to clean energy. A startup called Deep Fission of Berkeley, California, has raised $4 million in venture funds and is trying to develop a way to build nuclear reactors a mile underground. And, no surprise, AI data centers is a target use case. The company believes it can reduce the cost of building nuclear power plants

by bypassing the traditional need to encase reactors in 12-inch thick steel and to build containment buildings with 6-feet thick concrete walls. Yeah, trading off one complexity with another that is a bit more practical is where a lot of novel ideas come from. If you're already burying nuclear waste deep in the ground, why not just go a bit deeper and do it all there? And if you manage to go a mile deep, then the earth will take care of some of what you have to build over ground.

However, construction and operations that deep in the ground is non-trivial. Besides staffing up, the pre-seed funding will also seek regulatory approval and flesh out all these issues. In a deal that underscores AMD's need to broaden its strategy for taking on NVIDIA's dominance in AI, AMD announced it will purchase data center and hyperscale system supplier ZT Systems. It's a move that obviously made sense given NVIDIA's proven ability to move from strength to strength with its GPUs, ARM-based CPUs, and also its growing HGX and DGX systems.

So the move has to be taken seriously because AMD has its own record of excellent execution. And it's a big $5 billion deal, though, where AMD and NVIDIA play, you could very well say $5 billion is not an enormous amount of money. You sure could say that. If you think a trillion dollar valuation or more is in the cards, then maybe a few billion here, a few billion there is not a bad price to pay. And the price continues to get more and more enticing. It was only in late July that Morgan Stanley projected over $200 billion in revenue just for NVIDIA, just for Blackwell RACs. So the message seems to be everyone needs to move downstream towards systems, data centers, clouds, and solutions. And if so, that's a

watershed moment in the whole supply chain from chip to cloud. I would say this trend towards systems started with Cisco more than a decade ago when they built their own servers, continued by NVIDIA, led to the acquisition of Mellanox by NVIDIA, and is now looking like it is being adopted by AMD. And that's going to pose strategic questions for everyone else. Well, on that note, that's it for this episode. Thank you all for being with us. HPC News Bites is a production of OrionX in association with InsideHPC.

Shaheen Khan and Doug Black host the show. Every episode is featured on InsideHPC.com and posted on OrionX.net. Thank you for listening.