@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20231023

Episode Date: October 23, 2023

- HPE Cray at Crusoe Flared Gas Data Centers - IBM NorthPole AI Chip - TSMC financials: AI, 3nm, Inventory - AI Frenzy, AGI, Brain Waves as Input - Exascale, Aurora, TOP500 [audio mp3="https://orionx.net/wp-content/uploads/2023/10/HPCNB_20231023.mp3"][/audio] The post HPC News Bytes – 20231023 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. Hi, Shaheen. This week, we'll start off with a quick item. HPE said its HPE Cray XD supercomputers were selected by Crusoe, a builder of modular

data centers to power their new cloud services. Crusoe will leverage the custom systems for large-scale AI models. But an interesting thing here, Shaheen, is the energy source. Crusoe uses wasted, stranded, or clean energy, such as natural gas that is otherwise flared, for its data centers? Yes, two aspects of that story look important. The source of energy, as you mentioned, which has also been pursued by crypto mining, which has a very direct financial incentive to use low-cost and renewable energy, and flared gas makes great sense there. The second is the word custom, which we have talked about for a couple of years. The idea that Postmore's law requires customization

and architectural approaches to performance from chiplets to exascale clusters, and we are seeing an example of that. You remember that IBM Research was pursuing a neuromorphic chip called TrueNorth. That approach is motivated by the human brain's ability to do some things very fast with very high efficiency. It essentially uses physical twins of interacting neurons. This week, IBM revealed a chip called North Pole, which combines

several approaches specifically for power-efficient AI inference, and it includes actual calculations, not just neurons firing pulses. The thing to watch is the growing need to have memory and computation in the same place, what we call in-situ processing. In a blog on this chip from IBM, they talked about a Cambrian explosion in AI. You and I have analysis of new AI chips coming up with Carl Freund of Cambrian Research. IBM has been a computing patent powerhouse over the last decade plus, including its power processors. So it'll be interesting to see if this gains more market traction. TSMC Shaheen reported year-over-year Q3 revenue decrease of 10.8%, while net income fell 25%, which, with the surrounding generative AI driven by GPUs mostly manufactured by TSMC, seems somewhat surprising,

given their strength in advanced chips. I see three factors here. One, demand for anything related to AI is strong, and everything else is flat or down. Two, demand for three nanometer and five nanometer chips is strong. And three, so-called inventory digestion, where customers are okay with lower inventories based on anticipated demand. You add them all up and it fits the curve. High-end AI saving the day for those who can align with it quickly. That's TSMC and NVIDIA foremost, and then others who can align with those two. Speaking of AI, there's a steady stream of AI news about artificial general intelligence, AI in robots, and sensors implanted in AI

in things and humans, even detecting brainwaves as input to AI. All of this indicative of the frenzy of development in a brand new, well-funded, fertile area. A really interesting thing here is the notion that if you think it, your thought can be input to AI, either via implants like Neuralink is pursuing, or even better by detecting brainwaves, as you mentioned, which is what Meta talked about this week. Baidu released their AI model that they said would rival ChatGPT4. OpenAI is more explicitly focused on AGI,

and the use of AI in warfare is a growing concern. Also, Shaheen, Exascale Day, having come around on October 18th, we did an update on the Aurora installation at Argonne National Lab. This is the Intel CPU GPU powered system expected to reach two exaflops of compute power. The team there is working feverishly on the installation and they're optimistic about impressive performance on the Ponte Vecchio GPU. But some questions remain, may remain, whether they will be on the top 500 list at SC next month. We'll see. Well, these systems face entirely new challenges

as they blaze new trails. Aurora is an enormous machine, 166 cabinets, 10,624 nodes, 300 miles of network cables, and the application and software stack running on all of it. We talk about the journey to exascale in our current episode of the AddHPC podcast with Paul Messina and the insights that drive this kind of project. It's not about meeting a deadline, but building the expertise. That's the real advantage that fuels national competitiveness. All right, that's it for this episode. Thanks 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.