@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20250414

Episode Date: April 14, 2025

- Argonne's nuclear reactor digital twin helps monitor, manage, train - An AI factory in the sky? Building data centers in space or on the moon using Space Based Solar Power (SSP) - TSMC may face US$1B penalty - Chinese AI players order $16B of Nvidia H20s - Hyperion Research says HPC-AI market grew a whopping 23.5% in 2024, poised to exceed $100B by 2028 [audio mp3="https://orionx.net/wp-content/uploads/2025/04/HPCNB_20250414.mp3"][/audio] The post HPC News Bytes – 20250414 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. Along with the HPC industry's permanent appetite for more compute power, over the last few years has emerged a permanent worry about having enough power

to supply the gigantic data centers serving the HPC AI infrastructure. With that comes concerns about greenhouse emissions from power plants that run on fossil fuels. That has revitalized interest in nuclear power, a clean, if somewhat risky, power source. We've talked about nuclear power over the past year or so, including, for example, Microsoft's effort to bring back into operations one of the reactors at the Three Mile Island facility in Pennsylvania, a site of a nuclear accident in 1979.

We also, only half jokingly, have said that HPC would be used to help perfect the use of small or large nuclear power sources. Argonne National Lab, near Chicago, was written up in the Wall Street Journal last week for its AI-based tool that can help design and run nuclear reactors at a time, as the journal said, when AI itself

is feeding a power frenzy. Argon's tool, the Parameter-Free Reasoning Operator for Automated Identification and Diagnosis, or PRO-AID, marks a technological leap in the nuclear reactor industry, according to the journal. PRO-AID is for new reactors, but it's also eyeing older nuclear facilities, many of which, like the one at Three Mile Island, are suddenly interesting and eligible to be resurrected. Yeah, ProAID creates a digital twin and then monitors and uses AI reasoning to diagnose and issue alerts when something may not be operating properly. It's a great example of HPC and AI coming together

and points to how these complex systems need to blend physical reality with physics-based modeling and agentic AI automation. The tool also has training benefits, since it retains institutional memory and best practices that new and existing employees can tap to improve their skills. So kudos to Argonne.

There were two pieces of news last week in the all-important area of advanced chips and geopolitics. Reuters quoted two sources who said that TSMC is under investigation by the U.S. Department of Commerce for violating export control regulations and, if found guilty, could face a $1 billion plus penalty. Allegedly, chips manufactured by TSMC for China-based design company Softgo have ended up in Huawei's advanced Ascend 910B AI processor. Huawei has been on the US trade restriction list for a number of years. TSMC is in such a strong position now that it'll be interesting how this transpires.

There's another report that semiconductors, cell phones, and other electronics have been given a much broader exemption from US tariffs. This exemption includes NVIDIA's H20 GPU, which was announced in 2023 specifically to meet then-current export restrictions to China. It's a lower performing version of the popular H100 accelerator. When the DeepSeq news indicated that lower performing GPUs may just do the job, Chinese demand for the chip shot up as local AI players like Tencent, Alibaba, or ByteDance all reportedly flocked to the chip, putting their combined orders at about 16 billion dollars just for the first three months of the year.

Growing challenges with AI data centers are causing some companies to go beyond thinking outside the box and think outside the planet. Imagine if you could have a data center on the moon or in Earth orbit and suddenly the big problems of power and cooling are eased. And since you trade in data, sending it back and forth is possible. That's what a few companies are pursuing now. Lone Star Data, which recently tested a miniature data center

on the moon via a mission by intuitive machines and launched by SpaceX. And Star Cloud is developing data centers in Earth orbit, initially at megawatt scale and envisioned to scale to gigawatt capacity. Of course, computers in space have been a long time coming. We remember HPE launching a system to the International Space Station in 2017. That system was called Spaceborne Compute 1, SBC 1, and was a proof of concept to see if commercial, off-the-shelf systems could operate in space for at least a year. It came back after almost two years. They followed that with SBC 2 in 2021, and that one came back in 2023.

And an upgraded version was launched over a year ago. The main motivation there was really edge computing, having computing power on-prem on the space station primarily for its own use. The main issue was radiation in space that is much stronger than on Earth, where the atmosphere weakens it. And the projects have shown that unmodified off-the-shelf hardware could work well in space for long periods, albeit with some software help to manage radiation. This time, the motivation is to build AI factories in space for use by anyone. Space-based solar power, SSP, can generate a lot more energy

because sunlight is more intense, again, because the Earth's atmosphere doesn't interfere, and you can position it so it's always in the sun and avoids any night times. According to the European Space Agency, a solar panel in space can generate 40 times more energy annually than a similar panel on Earth. A data center in space would also be easier to cool, since it's so cold out there anyway.

In fact, with vacuum and low temperature right there, maybe it will simplify quantum computing too. And it's easier to get it through regulatory hurdles since there is no local community impact and security is better since the system is physically isolated and data traffic can avoid a lot of terrestrial hops. And it's also fast given all the satellites orbiting the Earth. In fact, communication satellites of the future can have their own data centers,

like edge data centers do on Earth. Add to that shrinking system size and weight for the same performance, declining costs of launching into space, and soaring demand for AI factories, currently estimated at something like 20% compound annual growth for the next five to six years. And hey, this thing looks like a slam dunk. Are there any obstacles? Well, yes, there are a few. First, it all has to work and many things can go wrong. Then, launch costs are high, especially for larger cargo or assembly in space. There will be limited if no ability to repair systems.

Cooling requires specialized systems, since there is much less gravity and no atmosphere. An object in space already vulnerable to attack by adversaries, and highly-priced AI factories in space could escalate that. Interesting to watch. What industry analyst firm Hyperion Research now calls the HPC AI technical computing market

has traditionally shown good growth with the occasional blip downward. But in recent years, with AI combined more tightly with HPC than ever before, the industry's growth is increasingly accelerating. Hyperion announced last Tuesday that the market for on-premises server storage, software, and maintenance, plus spending in the cloud, reached $60 billion in 2024. This is 23.5% year-over-year growth and the highest annual growth they've reported in decades. They also said AI is now used by more than three-quarters of all HPC sites around the

world. Hyperion is projecting the overall HPC technical computing market will exceed $100 billion by 2028. Not just an Austin Powers-y kind of a number. Cloud usage would slowly grow to reach about 20% of that. All right, that's it for this episode. Thank you all for being with us. HPC News Bytes is a production of Orion X 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.