@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20250526

Episode Date: May 26, 2025

- Big AI in hot pursuit of AGI and SI - Stargate plans for US center with 400K GPUs, other sites in Mid-East and Asia - Would Nvidia invest in PsiQuantum? - Photonic and silicon-based Quantum Computing line up - D-Wave rolls out new Advantage 2 - RISC-V turns 15, positions itself "as a pillar of digital sovereignty on the world stage" - ISC25 June 10th in Hamburg with 195 exhibitors from 31 countries [audio mp3="https://orionx.net/wp-content/uploads/2025/05/HPCNB_20250526.mp3"][/audio] The post HPC News Bytes – 20250526 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. Big AI continues its pursuit of mega facilities. This past week saw news from OpenAI that it will build a $15 billion AI data center in Abilene, Texas that could house up to 400,000 NVIDIA Blackwell GPUs.

Company CEO Sam Altman said it will be the biggest AI model training facility in the world. The expanded project has secured $11.5 billion in new financing, according to a story in the Wall Street Journal, and has grown from an original plan calling for two buildings now to eight. More than $7 billion of that investment is reported to be coming from JPMorgan Chase. In addition, more details have emerged about a Stargate data center in the United Arab Emirates. Historian Data Center Dynamics reports that the massive campus in the UAE will

involve G42, OpenAI, Oracle, NVIDIA, SoftBank Group, and Cisco, and it will deliver 5 gigawatts of compute capacity. Additionally, a Stargate data center may be proposed for the APAC region as part of the $500 billion Stargate plan announced at the White House in January, soon after President Trump's inauguration. AI is a global trend that is just going to continue, not withstanding hiccups along the way.

AI needs GPUs, and it looks like it will continue to need GPUss and even more of them in the future with large fast memory and fast interconnects. And GPUs need a lot of power in a very small area and it looks like they will continue to need a lot of power and in an even smaller area. So AI's implementation will remain a question of securing, storing, and transmitting power and heat, power going into the data center and heat coming out and harnessed. Optimizing the overall picture in a holistic way is what is needed.

And GPUs have been in short supply, so you need to order them early and get allocation and go through an increasingly rigorous compliance process. How fast we can pull all of that together and also receive shipment varies in each case. 400,000 GPUs, for example, is a nice fraction of the annual supply of GPUs. For countries and very large companies,

the pursuit of general intelligence and then superintelligence is so strategic that whether or not AI is making money right now is probably secondary. You could say the same thing about industries like quantum computing, biotechnology, fusion energy, or cyber defense. Quantum made waves last week starting with D-Wave, and the announcement that its advantage too is now available. This caused a major share price jump for that stock. Going back a few months, NVIDIA CEO Jensen Huang famously poured cold water

on quantum exuberance when he said, the industry might be decades away from commercial viability. He later partially retracted those statements and at GTC, he had two separate sessions featuring quantum vendors on stage. He also announced quantum research centers and NVIDIA continues to build development software for quantum algorithms and emulation software running on GPUs. And they even took a minority

stake in SandboxAQ, a quantum software company spun out from Alphabet. All of this signals a move toward quantum and the growing market that it represents for GPUs. But that said, NVIDIA has not gone too far beyond research collaborations with quantum hardware vendors. So it's interesting and notable that according to the publication of the information, NVIDIA is in talks to invest in PsiQuantum, which is pursuing quantum processing units

based on photonics implemented in traditional silicon. Reportedly, PsiQuantum is in the midst of another financing round aiming to raise 750 million or more with a pre-investment valuation of six billion. Any investment is an endorsement, but would NVIDIA's expectations be limited to financial returns, or would it also imply possible inclusion into their product roadmap? We know Jensen's view of NVIDIA and its mission

that at its core it's all about driving compute power wherever and however faster compute power can be found. And of course, quantum has the potential to be the fastest compute imaginable. PsyQuantum is one of a half a dozen companies pursuing quantum computing based on photonics. Others are Orca in England, Xanadu in Canada, Quantella in France, Quicks in Holland, and Tundra in Wales. SciQuantum has raised about $2 billion if you include this last $750 million round. It is based in the US with Australian founders and secured massive funding from the Australian government and is building facilities in Australia. Its aim is to have 1 million photonics qubits on a single silicon chip.

There are other companies working on silicon-based QPUs, including EqualOne, based in Ireland and the US, Quantum Motion in England, Silicon Quantum Computing in Australia, and Photonic Inc in Canada. And Intel, which has a research project and occasionally shares very good progress. To refresh your memory, one way to look at different approaches to quantum computing is to consider that quantum effects are everywhere in nature.

So the question is how to harness them and manipulate them in a programmable way. So depending on the background of the scientists and engineers involved, they could work on photons, electrons, positrons, atoms, ions, or whole molecules. Modalities cover a wide range. Trapped ion, neutral atoms, cold atoms, photonics, silicon spin, quantum dot, topological, cat qubits, annealing, and they can be silicon based, superconducting, optics based, cetera. None has emerged as a leader,

and while a useful programmable quantum computer may not be decades away, it is likely about 10 years away. Special purpose systems like a physics lab apparatus that has some adjustable parameters, that could be earlier. Quantum communications would be even earlier than that, while quantum sensing is already here. So broad brush, that's the lineup right now. Quantum sensing or measurement, then quantum communications, then single application instruments, and then increasingly programmable systems.

One has to be really careful in interpreting the news in quantum computing. There are significant breakthroughs, but the space needs tens of them, not just a few. So the progress is real, but so is the hype. So if you need to make a big decision about quantum computing, you definitely should consult with those who track the technology closely. This month marks the 15th anniversary of the creation of the RISC-V CPU instruction set. It started at UC Berkeley, which has an impressive record of contributions to computer science and engineering. And as the name implies, it was the fifth such project at the university's famed Parallel

Computing Lab, or PAR Lab. RISC-V soon established itself as flexible and extensible architecture without red tape. A couple of years later, it got off the ground in a big way when it was launched with 42 founding members, which included vendors like NVIDIA, Qualcomm, Google, IBM, and Western Digital. In 2020, the governing organization of the technology called RISC-V International moved its incorporation and headquarters to Switzerland in what it said was a, quote, pursuit of unfettered access and geopolitical neutrality. They said this also included positioning the architecture as a pillar of digital sovereignty on the world stage.

Today, it continues to be the up and coming instruction set after x86 and ARM, and it's scaling its way up from controllers and embedded systems to desktop servers and GPUs. It's quite a success story. RISC-V has been very promising from the beginning. Several chips had attempted to be the open, scalable, extensible chip, notably Spark's

scalable processor architecture, which also has its roots at Berkeley and could go from sensor to supercomputer and was further developed and launched by Sun Microsystems in 1987. By 2010, the Berkeley team thought it was time for a fresh, clean start to incorporate the latest learnings

and to leave various baggage behind. So far, most of the RISC-V cores are in MCU's microcontroller units, small low-power silicon that is critically important but often buried deep inside chips. For example, in 2024, Nvidia shipped over a billion RISC-V cores in their chips. Each Nvidia chipset can include 10 to 40 RISC-V cores. There are several companies like Nvidia that switched from proprietary instruction sets to RISC-V often when they wanted to transition from 32 bits to 64 bits, when a disruption would have to be absorbed anyway.

Vendors building server chips based on RISC-V include Ventana, Sci-Fi, Tencent, Alibaba, Esperanto, Rivos, Andes, Semi Dynamics, and Microchip. In addition, the European Processor Initiative, EPI, is working on Victor and Tensor chiplets, or tiles, based on RISC-V. We're approaching ISC 2025 on Hamburg starting June 10th. And it's one of the biggest pure play supercomputing conferences of the year, though of course,

AI and other technologies are increasingly encroaching on traditional HPC. It's always a good event, rich in perspectives from European and Asian players along with the US. We hear from ISC folks that they expect 3,500 attendees and 195 exhibitors, a nice jump from last year, of which 47 are exhibiting for the first time. Exhibitors are from 31 countries.

This is in part due to increase in exhibits in AI and quantum computing. I expect there will also be a nice lineup of liquid cooling technologies. If you're unable to attend in person, they will have on-demand recordings starting June 13th. The URL is isc-hpc.com. Go check it out. 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.