@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20240401

Episode Date: April 1, 2024

- Microsoft OpenAI $100B Stargate supercomputer - Eviden winning deals, growing and hiring - Alibaba RISC-V datacenter chip - Linear Pluggable Optics Multi-Source Agreement, LPO-MSA - Nvidia banning CUDA translation software like ZLUDA [audio mp3="https://orionx.net/wp-content/uploads/2024/03/HPCNB_20240401.mp3"][/audio] The post HPC News Bytes – 20240401 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. Hi, Shaheen. You know, I've got to say, Shaheen, this AI stuff is really starting to catch on. We have a story from the information that Microsoft and OpenAI are planning a data center

costing as much as more than $100 billion and includes an AI supercomputer called Stargate set to launch in 2028. $100 billion is really astounding. It dwarfs, for example, funding for support of the U.S. microprocessor industry in the Chips and Science Act, which came in at $52 billion. Among other things, we wonder how this data center will be powered. We're reminded of AMD's Lisa Su's speculation last year that data centers of the future will be powered by their own nuclear power plants. Yeah, these numbers are really in a new territory. Talking

about $7 trillion for chips, $100 billion for a single data center, and then multiples of such large sites, that's a new mindset. If you see AI as a once-in-a-century opportunity, then yes, the stakes are high and a big spending is required to ensure future success. The counter argument, however, is that a rush into AI will cause too many failed projects. And to avoid a setback, the only way is to go slow and evaluate ROI on a case-by-case basis. Doug, you've been covering Eviden, whose systems get very good marks from customers for being at the leading edge. Eviden was carved out of Atos, its parent company, as the only European player in the space. But it looked like the financial engineering required to implement the carve-out was taking longer than anticipated.

Do I have that right? Yeah, we spoke with Emmanuel Leroux. He is Evident's head of advanced computing, HPC, Quantum, and AI, and he was emphatic in asserting that Atos's financial problems will not impact Eviden. He said they are winning deals, exceeding revenue goals, and they're hiring. And in any case, the French government issued a statement on March 19th that it will, if need be, back Eviden, which is instrumental to the country's goal of attaining what they call AI sovereignty. And in March, the company announced two sizable HPC AI deals in Europe.

Okay, a group of tech companies have announced an agreement to develop specifications for networking equipment and optical interconnects. The idea is to support an ecosystem of interoperable linear pluggable optics solutions, or LPO for short. These specs are intended to reduce the power, cost, and latency of moving data. The group's founding members include AMD, Arista, Broadcom, Cisco, Hisense, InnoLite, Intel, Macom, NVIDIA, and Semtech. Yeah, so the cool thing about it is that it's low power, does not incorporate a digital signal processing chip, a DSP, and it's protocol agnostic. The objective is to build switches, NICs, and endpoints.

The specifications will define the electrical and optical requirements for interoperability between multiple vendors of networking gear and optic modules, a bit reminiscent of the UltraEthernet consortium. Now, data center demand for networking is growing fast. Some estimates have it at about 10x over the next three to five years. There's news from Alibaba that the RISC-V R&D for a data center server chip is coming later this year. Alibaba was in the first wave of RISC-V supporters and a chip for datacenters had been mentioned before. It makes sense given the advanced chip technology export bans imposed on China in recent years. The story in the register said the CPU is built to power servers including those running AI workloads. Finally, a story on Tom's hardware said NVIDIA

has banned the use of translation layers for CUDA codes. This is as opposed to recompilation, which is at the source code level. The restriction has been in place since 2021, but gained visibility with a project called Zaluda, we covered it a few weeks ago, that puts a wrapper around CUDA binaries and translates them for a chip other than NVIDIA. So this could be another battleground in HPC AI that we'll watch and see how it develops. All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes 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.