@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20250602

Episode Date: June 2, 2025

- Nvidia Earnings - US Senators react to Nvidia’s plans for facility in China - New compliant chips for China by Nvidia and AMD - Trade restrictions reach EDA software - Vertical integration reaches China: Sugon-Hygon merger - AMD-Sanmina split ZT Systems - DOE NERSC’s “Doudna” supercomputer - UK-EU dial up supercomputer re-collaboration [audio mp3="https://orionx.net/wp-content/uploads/2025/06/HPCNB_20250602.mp3"][/audio] The post HPC News Bytes – 20250602 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 InsideHPC, and with me is Shaheen Khan of OrionX.net. As is so often the case lately, the intensifying geopolitical struggle between the US and China is dominating the world of supercomputing and big AI. China is an issue with the power to unite US politicians of all stripes. And last week there was the unusual, you could say unheard of, alignment

between two senators, one on the progressive left and the other on the MAGA right, who raised alarms about Nvidia's plans to open an AI center in Shanghai. Nvidia, for its part, said it was just a new lease for existing staff and was not a new initiative. Nevertheless, it served as a reminder of the geopolitical implications of Nvidia's presence in China. A Wall Street Journal article summed up U.S. worries in its headline, saying NVIDIA has said, quote, no advanced chip designs will be sent to the facility, and NVIDIA has been consistent about the value of getting AI developers, regardless of country, to build applications on a U.S. development platform, namely NVIDIA's CUDA. applications on a US development platform, namely, Nvidia's CUDA. Also last week, TechCrunch ran a story stating that Nvidia and

AMD are working on pared down GPUs that are allowable under current trade restrictions. It's a conundrum for Washington policy makers who want to increase exports to China and want to support American leadership and AI, but also don't want to give America's top geopolitical adversary access to powerful AI processors. And we should remember what our recent guest on the At HPC podcast, Dr. Handel-Jones, said about chip export restrictions to China, that they have been very effective in holding back China's progress in AI. So how to draw a balance between these conflicting goals is a major challenge. Speaking of trade

restrictions, they have now been extended to include chip design software. Top vendors in electronic design automation, EDA software, are Cadence, Synopsys, Siemens EDA, which used to be called Mentor Graphics before Siemens in Germany acquired the company for $4.5 billion in 2017, and Ansys, which has its beginnings in structural analysis but has expanded into EDA and other areas. In fact, all these companies are growing and acquiring other companies

and generally trying to provide very comprehensive, integrated, multi-physics, engineering, design capabilities across all kinds of disciplines. They all said they will comply, but this is likely to impact Synopsys the most, since it did about a billion dollars of business in China in 2024, compared to about 500 million for Cadence and about 125 million for Ansys. Siemens does not break out its EDA revenues, but it's a big diversified

company and has had presence in China for something like 150 years. Getting back to chips, you've all seen Nvidia's earnings, which beat expectations, coming in at $44.06 billion for the quarter, despite the $5 billion write-off and saying that it missed out on $2.5 billion of business in China due to trade restrictions. It said it will do just under 46 billion dollars next quarter so their year-over-year comparisons are going to start sounding more normal. In the coverage of all these there was some pricing information about chips. Reuters article mentioned in video's chips for China based on the Blackwell architecture would be priced

in the $6,500 to $8,000 range compared to the H20 chips that would be higher performance and go for $10,000 to $12,000 each. That's a bit of a clue on how their high-end chips would go for and how it all adds up to the billions in revenues. Continuing with HBC AI's place in the struggle between the US and China, two major Chinese tech companies, server and supercomputer maker Sugon and chip company Higone, announced last week they will merge. And the result could be an integrated server and CPU maker capable of developing significant supercomputing power. Sougan, also known as Donning Information Industry Company, has used chips from Higon

and is no stranger to supercomputing. In the June 2020 top 500 list of the world's most powerful supercomputers, it had 68 systems in the top 100. Higon, also known as Tenjin Higon Advanced Technology Investment Company, got a license to AMD's first-generation Zen x86-64 architecture in 2016 and has produced a mid-range variety of such chips.

A story in the register said that Chinese media recently reported Haigon plans to unveil a CPU with 128 cores and capable of running 512 threads, four threads for each core, which is something Intel and AMD CPUs are not capable of. The US Bureau of Industry and Security in 2019 added both companies to its entity list of companies suspected of conducting activities contrary to US national security and foreign policy interests. Besides the geopolitical angle, the technology trend here is the rebirth of the vertically integrated systems company, where it's all built in-house.

You could say Nvidia ignited this trend with their DGX and HGX strategy where they design it but do not quite manufacture or sell it. You have to buy it from vendors in the Nvidia partner network who are expected to add some kind of value, say total solution integration and professional services. Then AMD joined the fray with the acquisition of ZT Systems for $4.9 billion. It was announced in August 2024 and closed earlier this year in March 2025 with the idea that AMD would keep ZT Systems Design

and engineering teams and divest its data center infrastructure manufacturing business. AMD just announced that it has done that, working out a $3 billion deal with Sanmina Corporation. Sanmina is a big electronics manufacturing services, EMS company, based in the US US with some 35,000 employees and about seven and a half billion dollars in revenue and offices in 21 countries. Sanmina's Viking Enterprise Solutions Group has its roots in a company called

New ISIS formed in 2000 to build systems based on the then new AMD Octoron chip. New ISIS also had very interesting CC-NUMA technology called HORUS, and then renamed ExtendD Scale, which could enable 32 processor shared memory systems, an architecture that is now very popular with high-end AI systems. Sanmina acquired NuISys in 2003 and renamed it Viking in 2018, when the division expanded to include storage. Right around the time Sanmina acquired Neuasis, Cray acquired Actiga Bay and Sun acquired Kealia, both of which were also focused on Opteron-based systems.

Vertical integration is how the computer industry started, with mainframes and mini computers which built everything from chips and disks to apps. The company NCR was a famous case study when the trend went the other way, and companies that specialized in a system component were providing much better technology than those who did the whole thing themselves. The trend started with peripherals and then the OS, and then moved to everything else. It fueled and benefited from open systems which used open standards to simplify integration. Those concepts continue to be there. Customers want integrated solutions

based on open standards, but the solutions are now bigger and more complex. So the piece parts, the proverbial Lego pieces, are now systems and racks versus chips and motherboards. Dell and Nvidia have won a contract for NERSC-10, which the Department of Energy said is the next flagship supercomputer at the National Energy Research Scientific Computing Center at Lawrence Berkeley National Lab. The system is called DUDNAA, named for Jennifer DUDNAA, the Berkeley lab scientist who won

the 2020 Nobel Prize for Chemistry. Nvidia said the new system will outperform its predecessor, Perlmutter, by more than 10x while only using 2 to 3 times the power, which translates into 3 to 5x increase in performance per watt. Shaheen, we recently mentioned DOE's next generation integrated research infrastructure, the IRI, which will link National Lab Supercomputing

resources into an integrated whole. And NERSC said DUTNA will be connected to DOE facilities and allow scientists to stream data into the system from all parts of the country and to analyze them in near real time. Looks like the time has come for AI to give a company more HPC credentials. It is an interesting move into the leadership class segment by Dell, which has traditionally shied away from the very high-end HPC business with only an occasional win, but generally letting

HPE and IBM and then Lenovo go for it. I see this as a direct impact of HPC enabling and then using AI, AI becoming mandatory for the general IT market, and all players needing to jump in. So that led to Dell getting much closer to NVIDIA to take AI factory supercomputers to the enterprise market. And if they are doing billions of dollars worth of AI factory supercomputers, they might as well cover large systems for traditional HPC. computers, they might as well cover large systems for traditional HPC. Now, the system is targeted for deployment in 2026 and will be based on NVIDIA's

next generation VARUBIN chips and InfiniBand Interconnect. That probably be later in 2026, given that Blackwell just went mainstream recently. OK, as we head toward the ISC conference next week in Germany, there's news from Europe that the UK is, you could say, de-Brexiting itself where HPC is concerned. The UK officially joined the European High Performance Computing Joint undertaking, EuroHPCJU, in May 2024, providing the UK with free access to Euro HPC resources. Furthering this effort, public sector organizations in the UK have been invited to host an AI factory antenna,

part of an EU wide effort to link British research expertise to its advanced supercomputers across Europe. You know, we used to talk about an information factory years ago. So kudos to Nvidia to make AI factory a thing in day-to-day vocabulary. We're some ways off from a real data supply chain, which is when it will really start looking like a factory. Which reminds me, the word factory is from the Latin word factorium, meaning place of

doers makers, presumably meaning actual people who make things. So we shall see how the meaning holds up. Anyway, it's good to see this and it will no doubt advance science. Alright, that's it for this episode. Thank you all for being with us. HPC News Bytes 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.com and posted on OrionX.net. Thank you for listening.