@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20241118

Episode Date: November 18, 2024

- Supercocmputing-24 conference starts today - TSMC, CHIPS Act, semiconductor demand - Sandia National Lab and Cerebras [audio mp3="https://orionx.net/wp-content/uploads/2024/11/HPCNB_20241118.mp3"][/audio] The post HPC News Bytes – 20241118 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. We're all excited for the annual Supercomputing Conference SC24, which officially starts later this evening here in Atlanta, which makes Atlanta the center of the HPC world for the next few days. As usual, there will be a new edition of the Top 500 list announced, and there

is excitement about potential additions toward the top of the list. We will have an episode of the At HPC podcast devoted to the top 500 and green 500 lists probably later today. We'll also be tracking a large number of announcements at this show, everything from novel chip architectures to the latest advances in storage, liquid cooling, and modular data centers, AI models, and on and on. The HPC community is an early adopter of technology. So what you see at this conference is, well, everything, because everything can make supercomputers go faster in some fashion or control energy consumption in some fashion, et cetera. Topics on my radar this year include interconnects. We discussed with Torsten

Heffler a few days ago, and be sure to listen to the latest episode of the AdHPC podcast for that discussion. Along with CXL and memory technologies in general, there's been a quantum village at the Supercomputing Conference for the past few years, so I'm also looking forward to catching up with the growing presence of HPC in quantum computing. We've talked about the CHIPS Act, and right now I believe it will continue strong, and I hope any adjustments or relaunch of it will not be disruptive. There was news this week that TSMC is set to receive an additional $5 billion in low-cost loans to boost capacity in the U.S. This is in addition to $6.6 billion in direct funding, and Intel is to receive $8.6 billion by comparison.

Yeah, it's interesting to note that these efforts need massive investments. The $6.6 billion complements some $65 billion that TSMC is investing itself. We should add Applied Materials to the list of companies to track. It is the largest equipment maker for chips in the U.S. News articles report that its high-end equipment continues to see strong demand in the U.S., driven by AI and factory build-outs in the U.S. Outside of the U.S., it's below expectations. We keep an eye on developments among the AI chip companies, and Cerebrus, with its dinner plate-sized wafer-scale engine processor, attracts a lot of attention. We've surmised that the national labs, which are always interested in the highest-end compute power and in new architectures, might adopt Cerebrus technology

in their next round of leadership supercomputers. Last week, Sandia might adopt Cerebrus technology in their next round of leadership supercomputers. Last week, Sandia Labs and Cerebrus unveiled a cluster composed of four Cerebrus CS3 systems to be used as a Sandia testbed to expand research into AI workloads for national security missions. This isn't surprising. As we've discussed, Cerebrus, Zambanova, and Grok are all used at Argon's AI testbed facility, given some of the eye-popping performance benchmarks Cerebrus has reported. The first four Cerebrus CS3 nodes of a planned eight-node system named Kingfisher were recently deployed at Sandia, funded by the National Nuclear Security Administration. Yeah, testbed seems to be the operative word here, since Sandia has announced

collaborations with several novel architectures, similar, as you mentioned, to Argonne's testbed. Sandia said one of their interests is large-scale trusted AI models on secure internal tri-lab data. Tri-lab refers to Sandia, Lawrence Livermore, and Los Alamos labs. Because Cerebrus is a single massive chip, it eliminates a lot of complexity that comes with coordinating and running AI models across many GPUs. And because all those cores and memories are close to each other physically, it typically has better power efficiency. We'll have to remind you of the specs because it's just fun. The chip is over 70 square inches with 4 trillion transistors that form 900,000 cores and 44 gigabytes of memory at 21 petabytes per second of memory bandwidth. It uses TSMC's five nanometer

process, so it's got a couple of more steps to go for future generations. There was also news that filings by OpenAI had disclosed that they had at one point considered acquiring Cerebris. Now, any company can say they considered buying another. So that part is an interesting. The interesting part is really the indication of OpenAI's aspirations beyond software. All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes 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.