@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20250331

Episode Date: March 31, 2025

- AMD MI355X to debut with a 30,000 GPU cluster at Oracle cloud - ASML to open repair hub in China - EuroHPC-JU's CINECA selects 140-Qubit Pasqal system - ORNL sees a path for quantum performance advantage for Fluid Dynamics - Classiq, Deloitte Tohmatsu, and Mitsubishi Chemical compress quantum circuits by 97% and 54% - Prof Torsten Hoefler of ETH wins ACM Prize in Computing [audio mp3="https://orionx.net/wp-content/uploads/2025/03/HPCNB_20250331.mp3"][/audio] The post HPC News Bytes – 20250331 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. Before the buzz had worn off from NVIDIA's big GTC conference in San Jose, AMD got some welcome and well-timed news a big win for its upcoming MI355X GPU accelerator, which is due this summer.

The TechRadar publication reported that AMD has won a multi-billion dollar contract to build a 30,000 GPU cluster for Oracle. The news was disclosed by Larry Ellison during the company's most recent quarterly earnings announcement. NVIDIA's GPU market share is estimated at 95%, but AMD's win could be an indication that it can carve out a foothold, perhaps a fertile foothold, in the AI accelerator landscape. The chips built on TSMC's 3 nanometer node and based on AMD's CDNA architecture are expected

to ship this summer, as you mentioned. And they will feature 288 gigabytes of HBM3E memory, that's high bandwidth memory, and that high bandwidth is up to 8 terabytes per second. And they support FP6 and FP4 low precision computing that's become more and more important for AI applications. There's lots of speculation about where AMD GPUs will fit in the AI compute market and their price performance sweet spot, if you will, which would allow AMD to land and expand, as they say.

The publication TechRadar also says the MI355X is positioned, quote, as a strong rival to NVIDIA's Blackwell B100 and B200. So it's right up there. We often hear AI industry figures say that welcome an alternative supplier to NVIDIA, but NVIDIA is not only ahead on AI compute, but also software and the whole ecosystem, and by a very large margin. So AMD has a big task ahead of it competitively.

The good news is the market is growing and other companies can have a lot of fun without having to beat NVIDIA. The news is also just as much about Oracle's ascendance in the public cloud market, driven by good technologies, lower cost messaging, and pulling their captive software customers into their cloud like Microsoft did. If there's one aspect of the advanced computing pie where the West has a decided advantage, it's in high-end lithography EUV technology. That's extreme ultraviolet lithography equipment for making advanced chips. There's only one company that makes these machines the Netherlands-based ASML, Europe's largest technology company. US trade policy prevents ASML from selling

EUV machines to China, and without them China is at a significant disadvantage in the chip arena, that and the ban on exports to China of Nvidia's most advanced GPUs. But now, despite pressure from Washington, there are reports that ASML will build a quote, reuse and repair unquote center in Beijing this year, China being ASML's second biggest customer. Right. ASML disclosed the plan for the Beijing facility in its annual report that was released earlier this month. But they also sell lower-end DUV, deep ultraviolet machines, and Chinese companies, as you said, are a big customer. So there must be a lot of those lower-end systems there that also need service and support.

So it's not so out of the ordinary for them to do this. And ASML said they already had a repair center in China and this new one will upgrade and replace the old one. It's not a new addition. But in our jittery global scene right now, where all norms are being challenged, it's hard to tell if there is or there is not more to something like this. Is this business as usual or some kind of a statement about the spirit of the American chip trade policy towards China?

We'll be watching for sure. There's been a steady stream of quantum news in March. I know it was a focus of your solo hosting of News Bites last week, Shaheen, and we have a few more items to highlight this week. One is a 13 million euro deal awarded to the neutral atom quantum company Pascale by the EuroHPC Joint Undertaking Project. This is for a 140 qubit quantum system to be hosted at the Cineca Supercomputing Center in Italy. The system is slated for installation this year, and it will be integrated into the Euro HPC pre-exascale system Leonardo, with the intent of enhancing hybrid quantum classical

computing workflows. There's also news from Oak Ridge National Lab with promising work towards quantum advantage for scientific applications. Yes, ORNL reported successful testing of a quantum computing approach to solving classical fluid dynamics problems. Fluid dynamics is one of the hardest problems to model and solve realistically,

and it shows up everywhere, from biological systems to geographical systems to supersonic jets. And physical experiments are hard and expensive, if not impossible. So researchers have to simulate and they have to make simplifying assumptions based on computational capability. In other words, they have to balance accuracy with speed or scalability.

ORNL looked for quantum performance advantage for a simplified model and managed to find ways to suppress errors and formulate the problem just so, and were able to show that a path towards quantum advantage exists. They were upfront that the results are only quote on paper, but as usual that's a necessary stage in progress. The work is also significant because until now quantum physics and quantum chemistry were the most likely areas to find quantum advantage because, well, they were already quantum and formulated that way. Longer term, computations relying on tensor mathematics seem to have a shot at being implemented

for quantum computers. Speaking of problem formulation, quantum programs are often represented as a quantum circuit, which is an intuitive way to keep track of what operations are applied to what qubits and how they interact. It's kind of like old school flow charts in the early days of traditional computing. Once you create the circuit, you can look at optimizing it by compressing it, which makes it faster and also more likely to withstand errors. So Classic, a company that provides a very nice quantum software development environment, said that its work in Japan with Deloitte Tomatsu Group and Mitsubishi Chemical on materials

research showed that they could compress a quantum circuit by 97% in one case and 54% in another. This is kind of reminiscent of vector and paralyzation in traditional computing. Shaheen, one of our really outstanding guests on the HBC podcast, Torsten Huffler, who is a professor of computer science at ETH Zurich, that's the Swiss Federal Institute of Technology, has been named the winner of Association for Computing Machinery's Prize in Computing for Fundamental Contributions to High Performance Computing and the AI Revolution. The prize comes with an award of $250,000

from an endowment from Infosys, the digital services and consulting firm, and it's one of the most prestigious awards in the technology industry. Big congratulations to Professor Heffler. The ACM Prize in Computing is a great recognition of the impact of his research and leadership as the rising star in HPC that he has been. If you haven't heard our conversations with Torsten, he appeared with us in episode 92

last November 15th and spoke in part on his theme of the age of computation. Going further back to episode 49, he and Professor Satoshi Matsuoka of Riken joined us to discuss their paper written together with other colleagues on 12 myths and legends in HPC. Congratulations again, we're delighted that this is happening. All right, 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.net.

Thank you for listening.