@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20240729

Episode Date: July 29, 2024

- Nvidia to build GPUs specially for the Chinese market to comply with export control - Morgan Stanley report projects 60,000 to 70,000 Nvidia AI racks in 2025 - China and TOP500, no new news - NTT and U. Tokyo use Graphene Plasmon in pursuit of faster opto-electronics [audio mp3="https://orionx.net/wp-content/uploads/2024/07/HPCNB_20240729.mp3"][/audio]   The post HPC News Bytes – 20240729 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 with Shaheen Khan. Under the heading, there must be many roads to so great a market, NVIDIA is reportedly developing a customized AI chip for China intended to navigate, not to say work around, U.S. AI export controls. According to an article in The Information,

NVIDIA plans to sell the chip with a special server designed to maximize the chip's performance. The chip, tentatively called B20, is part of the new Blackwell series. NVIDIA's top-of-the-line AI processor launched earlier this year. Expert control is nothing new in supercomputing, and it visibly caught up with NVIDIA in October 2022 when the A100 and H100 were banned. NVIDIA quickly came up with variants, the A800 and H800 GPUs, to pass through U.S. export filters. Also, kudos to NVIDIA for calling them 800, since the number 8 has associations with wealth, success, and status in the Chinese culture. The Chinese

conglomerate Inspire will reportedly partner with NVIDIA to build a server. The two companies have an existing partnership in China. Not surprisingly, there have been several such deals with Western companies over the decades, including, for example, a joint venture with Cisco announced in 2016. What we can be sure of is continued trade restrictions in high tech. Just a couple of weeks ago, we talked about a Goldman Sachs report that questioned the ROI of AI for organizations. This week, a Morgan Stanley report taking even more AI market share while generating more than $210 billion in revenues from its Blackwell servers in 2025 alone. For perspective, this would approach more than 10 times the revenue

AMD reported last year and four times that of Intel. The online tech site Tweaktown said Morgan Stanley expects NVIDIA to ship between 60,000 and 70,000 units of the Blackwell servers, costing two to three million each. NVIDIA is making the NVL72 and NVL36 GB200 AI servers, as well as B100 and B200 AI GPUs on their own. The Wall Street Journal put out a story last week about China's supercomputing capabilities. It was a roundup piece on China's, quote, newfound secrecy about its HPC capacity. But this, of course, is not something new to the HPC community. We've known this dating back to more than five years when China stopped participating in the biannual top 500 ranking of the world's most powerful supercomputers. So the story was not a shocker.

No, it wasn't. The journal interviewed HPC luminary Jack Dungara. Jack was a guest on this podcast. It was episode 23 in May of 2022, which is another must-listen episode. The quote in the article is something he had said before. Quote, the Chinese have machines that are faster, they just haven't submitted the results, end quote. And that can be concluded from scientific papers published by Chinese researchers.

One system mentioned in state media is Tianyi-3, while the other is a Sunway. Earlier versions of these systems are on the top 500 list, including some information on their architecture, so one can glean how they might have been improved. In general, these systems appear to be a lot less general purpose than the American variety, and China has smaller versions of those too. As to why this is happening, proposed theories are, one, China does not want to flaunt their advantage to keep trade negotiations calm, two, conceal how far they lag. Or the opposite. Three, conceal how far ahead they are. We all know and love graphene and buckyballs. Graphene being a single layer of carbon atoms in a super tight hexagonal shape. It's a nanosheet, one atom thick with high conductivity. Well, applying an electromagnetic field to graphene can create electromagnetic waves that

move through the graphene sheet and couple with electron oscillations. So the electromagnetic field can be a light source and the sheet can have a metallic substrate and the hybrid structure can improve up to electronic performance. And we know that because NTT and the University of Tokyo reported they have generated and controlled graphene plasmon wave packets with a pulse width of 1.2 picoseconds. That's 10 to the power of minus 12. Why is that cool? Because it enables faster, cheaper optics and electronics. And because it shows just how much progress can be realized. The research team in Japan says their work is a step toward ultra-high-speed signal processing and achieves a 35% conversion efficiency that's significantly better than previous work in this

area. This paves the way, they say, to future developments in THC signal processing, including advanced elements like variable frequency filters, amplifiers, and modulators, all potentially leading to integration of photonics with electronic technologies. All right, that's it for this episode. Thank you so much for being with us. HPC News Bytes is a production of OrionX in association with Inside HPC. Thank you so much for being with us.