@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20241104

Episode Date: November 4, 2024

- India rising - High Bandwidth Memory (HBM) in short supply - Novel accelerator architectures - CHIPS Act funds Extreme Ultra Violet Lithography technology in the US [audio mp3="https://orionx.net/wp-content/uploads/2024/11/HPCNB_20241104.mp3"][/audio] The post HPC News Bytes – 20241104 appeared first on OrionX.net.

Welcome to HPC News Bytes, a weekly show about important news in the world of supercomputing, AI, and other advanced technologies. Welcome to HPC News Bytes. This is Shaheen Khan with OrionX.net, and with me is Doug Black of InsideHPC. Our first story is about India, the rapidly re-emerging global power where every company is looking for growth and significant collaboration. So now it's the US, Europe, China, Japan, and also India. Of course, throughout history, India has been a global power, and we are seeing that happen again. India has a lot going for it. Excellent universities, scientists, a large and growing economy,

a powerful worldwide diaspora that includes world-class management talent in India and internationally, who actually also run a good portion of the high-tech industry in the U.S., and also harmonious relationships with other major powers, including the U.S., Europe, and China. This is commonly contrasted with China's challenges, with a declining population and contentious relationships with some neighbors and the Western Hemisphere, which has also led to trade sanctions. We'll be watching and tracking with interest. Yes, India's growing presence in the AI industry is significant. And recently, that country and NVIDIA announced major orders for NVIDIA chips. As reported by TechCrunch, NVIDIA announced partnerships with Indian firms

to deploy its chips at its AI Summit in Mumbai. Among the firms in India lining up for NVIDIA chips are Reliance Industries, Tech Mahindra, Tata Communications, and Yara Data Services. Jensen Wong of NVIDIA said the deals include orders to the scale of tens of thousands of H100s, specifically by Tata and Yara, while Reliance will deploy NVIDIA's Grace Blackwell 200 technology. The chronic shortage of GPU supply has been a major thorn in the side of organizations seeking to implement big generative AI compute infrastructures and for the vendors that sell them. But now that the GPU supply chain is improving, The Economist reported recently that high-performance memory chips could be the next AI bottleneck. Right now, SK Hynix, which placed an early bet on HBM, has more than 90% market share for HBM3 chips, with Samsung playing catch-up. This is an exploding market that's sized at $18 billion this year, up from $4 billion last year, and projected to hit $81 billion by 2026.

While there's a lot of AI washing going on with companies adding an AI veneer to what they were doing, HBM, high bandwidth memory, has emerged as a real requirement for AI applications. The players are Micron, Samsung, and SK Hynix, with a very interesting history of several mergers and restructurings that involves Hyundai, LG, and SK Telecom. Hynix is a mix of the words high and electronics. The Economist article notes that SK Hynix and Micron have pre-sold most of their HBM production for next year. Both are investing heavily in manufacturing capacity, but that takes time. SK Hynix is reportedly working to use TSMC to boost capacity, and Samsung is working to recover from a slight setback in its production.

A theme in our podcast is future supercomputing architectures, from chips to apps. CPUs are advancing with on-chip vector and matrix extensions, blurring the line between CPUs and GPUs. GPUs have added transformer engines, dynamic precision management, and bigger memory to handle large parameter models. FPGAs have always been good for rapid prototyping and for very long pipeline applications like streaming data, but programming them has been difficult and the compilation process, known as place and route, has been hard and slow, but FPGA companies are working to address those problems too. And then we have coarse-grained reconfigurable arrays, CGRAs, which provide fewer but bigger building blocks than FPGAs. They give up bit-level

programming of FPGAs in exchange for easier reconfiguration. The RIKEN Research Center in Japan and home of the Fugaku supercomputer has published important papers on CGRAs. In simple terms, they have a collection of switch blocks, processing blocks, and buffers and can rearrange data flow among them for different tasks. Then there have been efforts

to decompose applications into mathematical graphs and state machines, transform the graph, and match it to available hardware and interconnect. So what's next? Well, an intriguing technology was previewed this week that points to a more dynamic way of getting performance from hardware for applications. Yes, NexSilicon, whose novel compute architecture has been discussed for more than five years by the company and industry observers came out of unconcealed stealth last week, unveiling what it called the Intelligent

Compute Accelerator. The idea is to use telemetry to identify the likely flow of an application and optimize the chip's operations accordingly until the majority of the chip's compute resources are devoted to those operations. Their announcement included a positive quote from Sandia National Labs that they are using NexSilicon, quote, in anticipation of our deployment later this year of our new integrated system named Spectra to be delivered in 2025. High-end chip fabrication relies on extreme ultraviolet lithography, and boosting it in the U.S. is part of the CHIPS Act. Even as we hear rumors that one of the presidential candidates might kill the CHIPS Act if he is elected, the White House announced an $825

million effort to develop an American Extreme Ultraviolet lithography R&D center in Albany, New York. Shaheen, I may have given away which candidate wants to end the CHIPS Act by using the male pronoun. You may have. To refresh everyone's memories, extreme ultraviolet, EUV, is the current leader in high-end lithography, and the Dutch company ASML is the only game in town, selling multi-hundred million dollar school bus-sized machines that provide it. When ASML introduced it, it was billed as the machine that saved Moore's Law. So all high-end chips need it, from cell phones to high-end GPUs, manufactured by TSMC, Samsung, and Intel for now. The base EUV technology came out of the U.S. National Labs and was licensed only within the U.S., but one of the licensees, SVG, was acquired

by ASML, and they worked many more years to make it work, including close collaboration with TSMC. So re-establishing domestic competitiveness with EUV and what comes after that is a major cog in the CHIPS Act strategy, which is to reduce U.S. dependence on foreign suppliers of semiconductors. 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 Thank you for listening.