@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20250825

Episode Date: August 25, 2025

- FugakuNEXT, Japan's next leadership supercomputer led by RIKEN - Fujitsu MONAKA CPU - Fujitsu's 1000-qubit Superconducting Quantum Computer - CPU-GPU-QPU hybrid systems coming - US Government's equity stake in Intel: "Arsenal of Democracy" or "intervention"? - Softbank bets on Intel - New geopolitical football: H20 GPUs [audio mp3="https://orionx.net/wp-content/uploads/2025/08/HPCNB_20250825.mp3"][/audio] The post HPC News Bytes – 20250825 appeared first on OrionX.net.

Welcome to HPC Newsbytes, a weekly show about important news in the world of supercomputing, AI, and other advanced technologies. Hi, everyone, welcome to HBC Newsbytes. I'm Doug Black of InsideHPC, and with me is Shaheen Khan of OrionX.net. Let's start with a look at Japan's Next leadership supercomputer called Fugaku Next, which builds on Fugaku, the former number one-ranked supercomputer in the world housed at the Rican Research Institute. Like that system, Fugaku Next will be built by Fujitsu and will employ arm-based CPUs, but it also will be the first flagship Japanese system to have GPU accelerators from NVIDIA. The system will deliver AI performance exceeding 600 Xaflops in floating point eight precision

and is expected to become the world's first Zeta scale supercomputer according to Rican. They're aiming for the system to be up and running in 2030, which appears to us to be a plausible timetable. Now, Sheen, I know you admire the Fugaku architecture, but I'm curious about the implications of the new systems design and whether that has implications for the next round of leadership supercomputers from the U.S. By that, I mean, is there anything about Fugaku next that you think is surprising or intriguing? Well, we started covering Fugaku next when it was mentioned at the Salishan conference in April.

And then again in June, when Rican was selected to drive the project, no surprises there. Fujitsu first talked about the chip, Monaco chip in 2022, saying it will be a mainframe class chip that continues with arm instruction set, as you mentioned, and a large-scale super computing capability, but we'll also add special focus on energy efficiency and data center throughput capacity. They showed a mock-up at SC23 and then disclosed a lot more at Mobile World Congress in May of this year. The idea of being able to add GPUs to the system was always there and became a lot more real in 2024 when Fujitsu announced a strategic partnership with AMD to connect AMD GPUs to Monaca CPUs and to support the AMD Rockham GPU development software.

So the choice of Nvidia here is a surprise, but not a big one. since they've also been saying they want to keep things as standard and as open as possible. For example, they plan to have enough AI performance within the CPU itself to have air-cooled configurations, and they will use DDR5 memory instead of high bandwidth memory, which also helps with energy efficiency, and they rely on PCIE and CXL-3, which will enable wider connectivity. Rican said the hardware improvements they anticipate developing with Fujitsu and NVIDIA, along with the algorithmic and software innovations will, within the same approximately 40 megawatt

power envelope as Fugaku, will deliver up to a hundredfold increase in application performance, which is comparable to what they did with the K computer as they transitioned from that to Fugaku. In addition, by advancing AI for science, including AI-driven hypothesis generation and validation, automated code generation, which is rapidly coming our way, and the automation of physical experiments, the project will seek to accelerate the entire cycle of scientific discovery. I'd like to add, because it's fun, that monaca is a Japanese suite that is made of often elaborately designed wafers with sweet filling in between. The monica chip will have two nanometer core dyes or chiplets, five nanometer S-ram dyes, and I.O. dies. So it is very

much of a layered cake itself. Okay, so continuing with Fujitsu, the company announced it has started development of a superconducting quantum computer with more than 10,000 physical qubits with completion scheduled for 2030. They say the system will yield 250 logical qubits and will utilize Fujitsu's star architecture described as an early stage fault-tolerant quantum computing architecture developed by the company. Fujitsu aims to make practical quantum computing possible, a promise we've heard from many vendors, and they're targeting areas like materials science. So a lot is targeted for 2030, and the team there has a good record of execution. So we can imagine Fugaku next to become a hybrid supercomputer, including CPUs,

GPUs, and QPUs for quantum processing units. Anyway, as part of this, Fujitsu will participate in quantum development projects that also will involve RECAN. Key components of the project will include high-throughput, high-percision cubit manufacturing technology, chip-to-chip interconnect wiring and packaging technology that's needed for scaling to those numbers of cubits, high-density packaging and low-cost cubit control, which would address challenges such as cryogenic cooling and control plane complexity, and decoding technology for quantum error correction. Chips made geopolitical news again with the Trump administration announcing that U.S. government will take a 10% stake in Intel, a deal that marks,

as the Wall Street Journal reported, the latest in a series of extraordinary private sector interventions by President Trump. The agreement calls for nearly $9 billion in grants that were awarded to Intel from the 2022 Chips Act but not yet paid out to be used as equity, and the federal government getting more suck depending on how Intel's chip fab business fares in the future. It was earlier this month that Trump called for Intel CEO Lip Butan to resign for his alleged, ties to China. So there's been a significant turnaround in Intel government relations. Also, it's interesting to see other politicians' reaction to this transaction along with the White House deal with Nvidia and AMD earlier this month in which the companies will share with the government 15% of revenue of GPU sold to China. While Republicans typically oppose government

involvement in private industry and Democrats tend to support it, those roles have been reversed with these latest developments. As for funds for Intel, in addition to this $9 billion, SoftBank signed an agreement to invest $2 billion in Intel common stock at $23 per share. Reestablishing U.S. leadership and chip manufacturing is one of the core issues in national security, and Intel is the only short-to-mit-term option for the U.S. to have a shot at it. So while national security was not a front-and-center topic of news articles, these cash infusions will help Intel push ahead with his chip manufacturing modernization. I don't think they will be big enough, but they will help. As for the reappointment of Chips Act money to pay for equity in Intel,

the government will be a passive owner of non-voting stock with no governance or information rights, no board seats, nothing that could imply direct oversight or access to sensitive internal data, and for matters that require a vote even from holders of non-voting shares, the government will vote with Intel's board. This is in contrast to when U.S. steel was bought by the Japanese Nippon Steel, where the U.S. government asked for a board seat and veto powers. So perhaps the best way to see this is as a signal that the U.S. government is serious about Intel's success. That would be aligned with national security objectives. But questions about government intervention in private corporations will remain, and some observers would prefer that this be

a grant or even a loan rather than equity. Doug, you reminded me how during World War II, the U.S. government worked with private industry to retool our manufacturing base, enabling the U.S. to become, as was said at the time, quote, the arsenal of democracy. Ah, yes. Now sticking with geopolitics, now that NVIDIA convinced the Trump administration to allow sales of its H20 GPUs to China, in exchange for NVIDIA sharing 50% of the revenue generated by the chip, The Chinese government has decreed that the H20 poses a national security risk, resulting in NVIDIA-posing production of H-20s.

A story in the information said that NVIDIA has directed chip packaging firm Amcor and high-bandwidth memory supplier Samsung to stop production of the semiconductors. Interesting chess games of technopolitics, I guess. It was also reported that H-20's performance is on par with locally produced Chinese chips, and roughly a third to half the power of NVIDIA's flagship Blackwell GPUs. So this would imply that China does not need these chips. But we also just reported last week that the locally produced Huawei Ascend GPUs had experienced chronic problems running AI learning workloads,

which would imply that China does need American chips. So which is it? In any case, in response to these developments, the Wall Street Journal reports that NVIDIA will now seek approval to sell a more advanced AI chip to China, though still presumably significantly less capable than Blackwell. The journal said, quote, they are betting that the product performs better than chips

Chinese companies can offer, but worse than the most advanced in video products, and that both governments will be satisfied with the outcome. All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes is a production of Orion X in association with Inside HPC. Khan and Doug Black host the show. Every episode is featured on insidehpc.com and posted on

Orionx.net. Thank you for listening.