@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20240923

Episode Date: September 23, 2024

- AI-fueled M&A rumors: Intel, Qualcomm, Ampere - Three Mile Island resurrected to feed GPUs for Microsoft - Beyond GPUs lie specialized chiplets [audio mp3="https://orionx.net/wp-content/uploads/2024/09/HPCNB_20240923.mp3"][/audio] The post HPC News Bytes – 20240923 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. This is a story that only a few years ago would have been unthinkable. Intel, which some of us used to call Chipzilla, could be an acquisition target. There are reports that smartphone chip giant Qualcomm is targeting Intel, whose stock is worth about half of its 2021 value.

Of course, even if Intel is receptive to Qualcomm, a buyout would face regulatory approval. In fact, Intel's attempt to buy Tower Semiconductor for $5 billion came to an end last year when Chinese regulators objected. Intel recently announced plans to split off its founder business, and it plans to lay off 15,000 employees. So much of this, of course, is due to the big market shift to AI and servers powered by GPUs, where NVIDIA stole a years-long march. AI means big money, but only for those who are driving the industry. If you're among them, you have valuation to go shopping. If you're not among them, you could be a target. So this week, we witnessed more M&A rumors accompanied by what I consider narrow analysis that could justify it. Qualcomm presumably approached Intel to discuss a deal, and Ampere, the maker of merchant ARM CPUs, apparently sent out feelers to be acquired.

As you've heard me say, it is best for the nation and even the markets to leave Intel alone and let them complete the massive transformation towards which they have made excellent progress. That is, if you believe a broader semiconductor supply chain is a geopolitical imperative. Throughout the history of the advanced computing industry and of the industrial revolution, apparently insurmountable walls have been thrown in the way of progress until those walls come down. Today, we are up against daunting limits to the potential of HPC AI, energy consumption being a major one. An article in The Economist called The Breakthrough AI Needs focuses on the staggering energy requirements of AI workloads and what can be done about it. This points to two trends,

the search for cleaner energy, such as nuclear power, and we see news that the old Three Mile Island plant is reopening, and even a clean jet fuel under development produced by photosynthesis, which sounds exciting, assuming it has wider application than just air travel. And it points to specialized AI chips and software. No energy, no data center is the mantra. AI's energy crisis is real, but it's a small fraction of the total cost. And when you need so much energy in one place, you'll have to be right next to the power source. That's pushing data centers to look at all the new ways to locate data centers and sources of energy. The global competition to stay on the leading edge of AI makes it impractical to slow down

or coordinate things. So while there are many efforts to lower electricity requirements, and especially in AI inference, there's also a push to secure megawatts and gigawatts of power to feed the existing energy-hungry GPUs. We've talked about geothermal, silicon oxide fuel cells, nuclear reactors a mile deep in the ground, small modular reactors, and even fusion, and several companies are working on these paths. This one, the Three Mile Island deal, is interesting since it's bringing a traditional reactor out of retirement for a single customer, and it includes very big dollars. Three Mile Island was the site of a partial nuclear meltdown, the worst nuclear accident in the US, and did a lot to dampen nuclear power

in the US. It is being renamed Crane Clean Energy Center with $1.6 billion to get the 880 megawatt reactor going and a contract to sell all the output to Microsoft for the next 20 years. It's big. There are other efforts to revive old reactors, also with big dollars behind them in Michigan and Iowa. And just a couple of months ago, Bill Gates put a billion dollars into a new nuclear power plant project in Wyoming. The Economist article that we talked about echoed in comments made by AMD CEO Lisa Su in an article in the Wall Street Journal. She conceded that GPUs are now the architecture of choice for large language models because they're efficient for parallel processing, but they, quote, only give you just a little bit

of programmability. Do I believe that that's going to be the architecture of choice in five plus years? I think it will change, she said. What Sue expects isn't a shift away from GPUs so much as a broadening beyond them. Though beyond that, she declined to be specific. As did The Economist, the journal cited Amazon and Google developing their own AI chips built for training and inferencing functions. Broadcom's CEO said this year that their custom chip division, which mostly helped Google make AI chips, is generating more than $4 billion in revenue. Yeah, the post-Moore's law period we're in, combined with AI as a new ecosystem, the edge-to-exascale nature of deployments, and increasing transistor counts lead to a need for architectural approaches to performance. That's why we have the Cambrian

explosion that we see in AI chips, so many of them targeting specific but big use cases. More money has flowed into funding AI chip startups in the first half of this year than in the past three years combined. And the past three years were big too, so it truly is a stampede. But developers keep making changes to AI software. Bigger models are sharing the spotlight with smaller ones and more specialized systems. For example, OpenAI's newest model, O1, tries to be better at reasoning versus generating text. Other makers are employing less demanding computations, etc., etc. So it's still a moving target. 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 you all for being with us.