@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20250113

Episode Date: January 13, 2025

- China's TOP100 - AI PCs, workstations, workgroup servers - DataCenter Capacity Growth - MPI ABI [audio mp3="https://orionx.net/wp-content/uploads/2025/01/HPCNB_20250113.mp3"][/audio] The post HPC News Bytes – 20250113 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. You know, it's possible this isn't generally known, but since 2002, the People's Republic of China has issued a listing of the country's top 50 and then the world's most powerful systems twice annually, with new lists released at SC and ISC. China's Top 100 list is issued by the Chinese Society of Computer Science. They recently updated it, and for China HPC watchers, it's a disappointment.

I saw the nice coverage on this on InsideHPC. The site is mostly in Chinese, but translation is not so hard these days, so we can look at some things. The list is not complete. It's missing real high-end systems. And as you said, there's very little data in the site. No mention of newer, higher power systems that were referenced in Gordon Bell Prize submissions. No configuration information on what is listed,

and a lot of generic info, like the vendor being listed as, quote, server provider. So yes, so much about China's top 100 list is opaque, and I continue to wish they would resume participating in top 500, even and especially as that project evolves. The top system listed has 160,000 CPUs and a peak LIMPAC score of 620 petaflops in 64 bits, which would rank the system as fifth

on the current top 500 list. A longtime friend of ours in the HPC community, Matt Walters of the OmniScale Media Public Relations firm, drew a nice line of analysis regarding the new Blackwell-powered AI workstation announced by NVIDIA at CES last week, noting that in 2010, the world's fastest computer was Cray Inc.'s Jaguar at Oak Ridge National Lab with a Linpack mark of 1.76 petaflops. Matt said that two years earlier, IBM's Roadrunner had made history as the first system to break the petaflop barrier, a $120 million, 250-ton supercomputing monster of its day. of AI performance for $3,000 in a desktop unit. Said Matt, what used to require government funding in a dedicated facility now costs less than a used car and fits on your desktop next to your coffee

mug. Well, you know, hail Moore's law. NVIDIA joined the AI PC movement when it announced Project Digits. It's a very nice, clean, and physically small box, and it's sure to fuel on-prem AI development. It is billed as a personal AI supercomputer and uses a low-end chip branded as GB10 drawn from the Grace Blackwell architecture, and runs a variant of Linux provided by NVIDIA. It's one petaflop of AI performance, by the way, which in this case means 4-bit precision. So comparisons with big systems and their 64-bit performance are interesting and directionally valid, but probably should be avoided. It doesn't run Windows like other systems being sold under the AI PC label or Apple Mac OS systems that are also common developer

systems and quite powerful for AI in their own right. In principle, all developer platforms should have easy access to NVIDIA software and cloud capabilities. So it really joins the fray of other systems. At the high end, and presumably for more money than the starting 3K price, you can get 128 gigabytes of memory, four terabytes of NVMe storage, and run a 200 billion parameter model. They also talked about pairing two of them with fast networking to run even larger models. All of that also make it look like a workgroup server to be shared by several developers. Okay. New data and forecasts from the

analyst firm called the Synergy Research Group has concluded that the average capacity of hyperscale data centers that will be opened over the next four years will be almost double that of current operational hyperscale data centers. No surprise, the explosion in hyperscale load is driven by generative AI, which are power hungry and have supercharged that trend. There will also be some degree of retrofitting of existing data centers to boost their capacity. The overall result is that the total capacity of all operational hyperscale data centers will grow almost threefold by 2030. The report is based on 19 of the world's major cloud and internet service provider firms. That includes the usual suspects, as well as large operators in software as a service, platform as a service, search, social media, e-commerce,

and gaming. The report says that by late 2024, those companies had 1,103 major data centers in operation around the world. When they look at capacity growth and new data centers, it shows, not surprisingly, a hockey stick growth starting in early 2023, a few weeks after ChatGPT ignited the generative AI rocket ship that we've been seeing since. Given the stampede to infuse everything with AI, you might say that tripling capacity by 2030 looks conservative, but cloud providers have the means and the agility to make it go faster or slower. The big risk, as usual, is upstream and high-end chip manufacturing, where too little or even too much demand is not

pleasant. A lot of investment has already gone into new chip factories, so the hope is that demand would not soften or spike even in relative terms. We'll close with a quick update on MPI, the communication library that originated in HPC and has become an industry standard. It has traditionally been an API, application programming interface, managed through subroutine library calls, which is a compile time process. But there has been an effort to provide it also as an ABI, application binary interface, which is a runtime thing. A lot of excellent work has gone into it, though some work remain on Fortran since its bindings are varied and need work. This is driven in part by the advent of containers and the desire to be more dynamic in selection of libraries. The news is that the MPI-ABI has been voted into the

next version of the standard that will be MPI 5.0 and expected to be released in June this year. 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 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.