@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20250203

Episode Date: February 3, 2025

- DeepSeek, 1+5 lessons - Rack Scale Arch redux, Intel GPU roadmap change - LLNL and OpenAI, national security apps, fully on-prem? - Google and Kairos Power, Small Modular (nuclear) Reactors [audio mp3="https://orionx.net/wp-content/uploads/2025/02/HPCNB_20250203.mp3"][/audio] The post HPC News Bytes – 20250203 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. Shaheen, if this podcast lasted an hour, it wouldn't be long enough for the DeepSeek bomb that went off early last week. And it's changing by the hour with impacts and implications emerging all the time as we learn more about it. But I came across a comment from the UK blog site

MetaWriter that captures some interesting points, including this, quote, the DeepSeek team built an amazing AI, even if the extent of the capital invested and the computing resources remain open to debate. More than a technical victory, it's a moral one. The project is truly open source, can be deployed natively on a $6,000 PC, and as such, is proliferating across the world. The quote goes on to say, Silicon Valley's brute force GPU model was not only inelegant, but clearly intended to concentrate massive computing resources in the hands of the business community and the United States. The implications of deep seek are geopolitical, economic, but also moral. So to be sure, I don't agree with everything this

writer said, including a later comment that the PRC regime is mostly benign. And I'm not sure there was an intentional plan to concentrate power in the hands of the US, AI power, that is. It's more likely that that's just how things have evolved to this point. But finding greater cost efficiencies, if that's what DeepSeek has done, is what people in the US like to call the American way. For now, before we fully understand DeepSeek, all we can say is that its challenge to establish AI business models that have attracted hundreds of billions in investment could be real. An explosive piece of news it was. We had covered it a few weeks ago, actually, but it went boom last week after benchmarks came out that showed it is superior to

its many competitors. The big lesson and challenge when news like this breaks is how to react to it. What if you're compelled to react right away, but it takes a few days or weeks to figure out what actually happened? So the first filter has to be what is the right sense of urgency? Are you overreacting or underreacting? What is your exposure to weaponized information and possible psyops? Are big conclusions invalid and too soon or valid and not soon enough? So as more verified information has emerged, we can start seeing an outline. One is open source. At first, it wasn't clear exactly what and how much of it had been open sourced. This seems to have been verified now by people who have set it up on local systems and made it work. Open source is not a new topic here with Meta as a big company champion. And in fact, DeepSeek leveraged the many open source

models that are out there on the Hugging Face site or others. Number two is lower costs. Open source means leverage and community and proliferation, but it also means lower costs. We heard a lot about the Jevons paradox and interpretation of the economic supply and demand model that says lower costs not only enable more supply, but also expand the market, leading to more consumption and growth for the total market. Number three, of course, is near and dear to our hearts. It's HPC. The performance optimization work that DeepSeq has done is pretty routine in the supercomputing world. Overlapping computation with communication, for example, goes back decades. Using lower precision numbers in more places is also an existing approach. It is possible it's related also to the background of the DeepSeek team,

coming from the so-called quants part of financial services, which often means scientific engineering background. Item four is quality. While release benchmarks beat the competition, we saw reports later in the week that DeepSeek has a much higher hallucination rate, including in the references that it provides. This implies that its effectiveness is within a narrow scope. Speaking of which, item five is just what the scope is. Is it domain specific, mixture of experts, or all the way to artificial general intelligence, which is to say that if DeepSeek has one-off chat

GPT, that may not be the ultimate goal of OpenAI and its ilk. The big US AI companies, whose approach is criticized by many, are after AGI. They're not trying to fix a problem for a specific use case. So while they probably have already caught up with the methods here, they will continue to have very expensive objectives with very fast cycles that can become even faster now. The durable aspects of DeepSeek's technical approach are very much HPC. So once again, all players would do well to take HPC seriously. DeepSeek was a curveball for chip and system vendors and perhaps an opportune moment for Intel to realign

its roadmap once again. The company canceled its Falcon Shores AI chip that was to follow Gaudi 3. It said it will use that internally as a test platform while focusing on the follow-on chip Jaguar Shores in a quote, system-level solution at rack scale. That would align it better with what NVIDIA is doing with Blackwell and DGX rack scale and DGX cloud. Look for Intel rack scale architecture online, and you will see a lot of news from 10 years ago. Look for Intel internet data center, and you will go back 25 to 30 years ago. Maybe there were good reasons to abandon those projects back then at the time, but it's hard not to see news like this as yet another setback imposed by the realities of the market and

after years of changing strategies. To be sure, Intel remains strong. It has a command of the traditional server market, which is not growing, but it is important and maybe things like DeepSeek give it more relevance to AI. It's strong in PCs and has very good software. And we sure hope its significant advances in chip manufacturing continue full speed. Interim co-CEO Michelle Johnston-Holthaus said she made the decision based on competitive feedback from customers and from Intel's own GPU engineering team.

Here's what she said. One of the things that we've learned from Gaudi is it's not enough to just deliver the silicon. We need to be able to deliver a complete rack scale solution. And that's what we're going to be able to do with Jaguar Shores. Los Alamos National Lab announced a partnership with OpenAI to run its models on the lab's NVIDIA-powered Venado AI supercomputer. This is an HPE Cray system unveiled last April that has 2,560 direct liquid-cooled Gracehopper superchips and is ranked

number 13 on the top 500 supercomputer list, coming in at just under 100 petaflops performance. Yeah, and naturally, Los Alamos didn't reveal much about how Venado will be used for national security-related workloads, other than to say OpenAI will install its latest O-Series reasoning models to conduct national security research. But this sounds like a fully on-prem implementation of OpenAI, and that makes it interesting too. Let's end with Google and its agreement with KROos Power that points to the growing adoption of nuclear energy to address the looming HPC AI data center power crisis. Last October, the two companies signed an agreement to deploy a U.S. fleet of nuclear

power plants totaling 500 megawatts by 2035. Last week, Kairos said it had made progress on an engineering test unit at its development campus in Albuquerque. This is all part of Google's 24-7 carbon-free energy and net zero strategy. Okay, that's it, folks. Thanks so much for being with us. HPC News Bites is a production of OrionX in association with InsideHPC. Shaheen Khan and Doug Black host the show. Every episode is featured on InsideHPC.com and posted on OrionX.net. Thank you for listening.