@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20241014

Episode Date: October 14, 2024

- AMD's new CPU, GPU, DPU, NIC - Foxconn's big supercomputer at Hon Hai Kaohsiung Center in Taiwan - AI drives Nobel Prizes in Physics and Chemistry - Yann LeCun's take on AGI [audio mp3="https://orionx.net/wp-content/uploads/2024/10/HPCNB_20241014.mp3"][/audio] The post HPC News Bytes – 20241014 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. I was in San Francisco last week at an AMD event where CEO Lisa Su unveiled a series of new CPUs, GPUs, and a DPU, the highly anticipated CPU in the Epic family, codenamed Turin, along with a new GPU in the Instinct family called the MI325X, and a data processing unit in the Pensando brand called Selina, and the AI Pro 300 series for

high-end laptops. The AMD launch coming right after the Intel announcements is another sign that maybe competition is starting to show up for NVIDIA in deep learning. And there's a very large and very different market that's just forming for AI inference. AI is hungry for big, fast memory, and MI325X ups the ante with 256 gigabytes of high bandwidth memory. And that lets AMD claim better performance than NVIDIA's H200 chip by 20 to 40 percent. These chips get announced a few times these days, and this one was already announced at Computex in June. So it figures that AMD teed up the next one, the MI355X GPU, coming out next year and sporting even bigger memory at 288 gigabytes. We should remember that

not all applications, even in HPC, can use GPUs, so the Torrent CPU announcement will be important too. AMD also announced a NIC, a network interface card, called Polara 400, which is a 400 gigabit Ethernet card, which AMD said is ready for the Ultra Ethernet Consortium specifications, UEC, but the UEC 1.0 specification is coming in early 2025. The card is highly programmable, so it can probably adapt as needed. Cloud providers use their own DPUs internally, but the general market for DPUs has been slow and forming. In separate news, Foxconn, the big Taiwanese manufacturer, is building a supercomputer in Taiwan at the Honhai Kaohsiung Supercomputing Center. Not surprisingly, this is

a project with NVIDIA and will use the Grace Blackwell GB200 chip with the 72-way NVLink interconnect. It's a very large system. Think of it as 64 of the new DGX racks NVIDIA announced back in March. With AMD buying ZT systems for cluster engineering talent, and now Foxconn building an exascale class system, not to mention NVIDIA, who started it by building full-on racks, we are seeing an important next generation segment finally emerging. Rack scale engineering and manufacturing, and the whole data center as a system that we've talked about for 20 plus years. Yes, the system is expected to exceed 90 exaflops of AI performance. Noting that the system is Blackwell powered, it's worth adding that neither Intel last month nor AMD last week compared their new GPUs to the Blackwell chip, underscoring that both companies

are in catch-up mode relative to NVIDIA. That said, the high demand for GPUs, along with NVIDIA's high profit margins, give both Intel and AMD an ample market opportunity with their new chips. AI for Science made big news last week when a so-called godfather of AI, Jeffrey Hinton, and professor and physicist John Hopfield were awarded the Nobel Prize in Physics for research begun in the 1980s drawing on physics concepts to invent artificial neural networks. And Google DeepMind CEO Demis Hassabis, DeepMind Director John Jumper, and University of Washington Professor David Baker jointly won the Chemistry Prize for their work on understanding the structure of proteins and for developing new proteins. The prizes have been most welcome in the AI community, but they have also

generated a lot of discussion on whether the awards are really for chemistry or physics, or is this a way to pull in computer science and redefine core branches of science in a more multidisciplinary fashion. I think this is also a celebration of HPC. Nature seeks optimal balance, and that can be formulated with mathematics, and matrix algebra makes it computable. So it seems understanding and predicting nature includes a lot of matrix multiplies, and that itself is great insight. Doug, you found an interesting story in the Wall Street Journal regarding cautionary comments on AI by Yann LeCun, the French-American AI celebrity who works at Meta Facebook and is also a professor at NYU. I believe AI can do a lot already, so just because it can't do this

or that, it doesn't mean it can't be disruptive. But the discussion here is exactly about whether AI is close to general intelligence and logic, and he says we are nowhere near that. Yes, Lacan's belief is that true AGI will require models that learn in a way that's analogous to how a baby animal does by building a world model from the visual information it takes in. Shaheen, I think we're in agreement that while there's validities to some of what Lacan is saying, we also shouldn't underestimate the tremendous power and impact generative AI is and will have. Trained on a specific task that it's suited for. Gen AI already has, in effect, AGI-level capabilities, and that will eliminate livelihoods and can be used by nefarious actors

to create great harm. But my gut tells me that Lacan is right, that it's likely we will need to develop a new AI system design with Gen AI playing a contributing role before AGI is achieved. Okay, that's it for this week. Thanks so much 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.