@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20230925

Episode Date: September 25, 2023

- Intel Gaudi2, Collaboration with Dell, Satability AI - Samba Nova SN40L, LLMs - Air Force Research Lab 12 PFlop System - Small Modular nuclear Reactors (SMRs) - CHIPS Act: DOD $238m award for Microelectronics Commons Regional Innovation Hubs [audio mp3="https://orionx.net/wp-content/uploads/2023/09/HPCNB_20230925.mp3"][/audio] The post HPC News Bytes – 20230925 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. Welcome to HPC News Bites. I'm Doug Black. Hi, Shaheen. So, question, is Intel in the AI supercomputer systems business, and is it a bona fide competitor with NVIDIA in GPUs? Intel announced last week it will collaborate with Dell on AI systems solutions. Separately,

CEO Pat Gelsinger said Intel Gaudi 2 GPUs and Xeon CPUs will power a system for Stability AI, an AI company in the UK. He said the system will be Europe's most powerful AI supercomputer. GPU shortage in the market is making it an open field, not just for the top three, NVIDIA, AMD, Intel, but also for the next tier or two. It's an unusual environment, intense competition, while demand outstrips supply. If you can ship it, you're in the game, but customers have to wait in line to receive it. Meanwhile, system vendors have a backlog of orders and some of them conditional to ability to ship within a window. So it makes sense to expose and streamline the supply chain from fab to app. While we're on AI, LLMs, large language models, are where the

revenue is right now, and software is an important advantage for NVIDIA and a barrier for the rest. So there's a growing trend to specifically target LLMs and to highlight available software, often aided by open source projects. Basically, vendors want to say, you can have the chip and use it too. Along these lines, Samba Nova announced a new processor, the SN40L, manufactured by TSMC for LLMs with their software suite. A 5 trillion parameter LLM can be supported on a single system node. This is about 5x the high end we typically hear. Yeah, Samba Nova, an AI specialty chip company, acknowledged the trend towards smaller models,

but they asserted that bigger is still better and that bigger models will become more modular. They said customers are requesting LLMs with a trillion parameters like GPT-4, but they also want a model fine-tuned on their data. Also, Shaheen, the Department of Defense has awarded $238 million in CHIPS and Science Act funding for eight of what they are calling Microelectronics Commons Regional Innovation Hubs. This is the largest award yet under the CHIPS Act, passed last August, and provides about $280 billion for research and manufacturing of semiconductors in the U.S. Excellent infusion. This is a $2 billion funding over five years. DOD said the Microelectronics Commons program will leverage these hubs for domestic hardware prototyping and lab-to-fab

transition of semiconductor technologies. In all, more than 360 organizations from more than 30 states will be participating in the Commons. In other news, the Air Force Research Lab has installed a 12-petaflop system, a good reminder that supercomputers are proliferating and doing heavy lifting. Yeah, the announcement was a little bit unclear. The system could possibly be part of a huge deal won two years ago by Penguin Solutions, $68 million contract covering several new HBC systems for DoD. The announcement, as I say, was light on details for system specs

or the intended use of the system, so there's not much more we can share here. Now, Shane, we've talked on this podcast about the use of small modular nuclear reactors, SMRs, for data centers and supercomputing centers, and how actually supercomputers would be used to design their own power supplies. And there's news on this front here. Yeah, a lot going on in nuclear power, though still too early. First, as a reference, the top supercomputer frontier is a 21 megawatt system. News is, Last Energy has customers now for 20 megawatt SMRs in the UK. New Scale got approval for 50 to 77 megawatt SMRs in the US this year. Oklo is planning 15 megawatt SMRs.

And Rolls-Royce is talking about 470 megawatt SMRs in 2030. And TerraPower is out there too. 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 Inside HPC. Shaheen Khan and Doug Black host the show. Every episode is featured on InsideHPC.com and posted on OrionX.net. Thank you for listening.