@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20240617

Episode Date: June 17, 2024

- New paper on "Managing extreme AI risks amid rapid progress" with an all-star author list - Investments in Taiwan, Chip War - AI Chip Landscape, Specialty AI Chips gain traction - Where would AI datacenters fin energy? Fusion, Geothermal, Hydrogen Fuel Cells - Apple Silicon in Datacenter? [audio mp3="https://orionx.net/wp-content/uploads/2024/06/HPCNB_20240617.mp3"][/audio] The post HPC News Bytes – 20240617 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. Let's start with an article that Shaheen picked up on in the journal Science that sounds the alarm about the potential perils of AI.

Quote from the article, increases in capabilities and autonomy may soon massively amplify AI's impact with risks that include large-scale social harms, malicious uses, and an irreversible loss of human control over autonomous AI systems. Then the authors say, present governance initiatives lack the mechanisms and institutions to prevent misuse. The article presents a plan that combines technical R&D with proactive governance mechanisms. Now, of course, this topic has been discussed for many years, and it's interesting to contrast it with another big AI topic that's gaining prominence over the last month or two, which is the possibility that AI is losing its steam, that AI is being integrated into many other technologies, and that soon it will become

somewhat invisible, less a phenomenon in its own right, and more a part of other things. By that, I mean there's a strange dichotomy between the two points of discussion. But Shaheen, I know you're impressed by this article and by the authors who put it together. What are your thoughts? The big challenge of AI is to maximize all the good it can do while preventing disruptions or worse. It is extremely important, global, and difficult. So you pay attention when 25 of the most celebrated AI researchers, social scientists, and philosophers come together to publish a document called Managing Extreme AI Risks Amid Rapid Progress. The author list includes names like Yoshua Bengio, Jeffrey Hinton, Yuval Noah Harari,

Daniel Kahneman, and it just keeps going like that. They propose a plan and advocate policies and funding that can help reorient technical R&D and establish a series of governance measures. Topics include oversight and honesty, robustness, interpretability and transparency, inclusive AI development, addressing emerging AI challenges like an AI system taking over safety measures, evaluation for dangerous capabilities, evaluating AI alignment, going beyond behavioral evaluations to make sure AI doesn't behave well only when it is being evaluated, for example. Risk assessment, resilience, and then governance measures like institutions to govern the rapidly

moving frontiers of AI, government insight and awareness of AI development and its risks, moving away from the assumption that AI is safe unless proven unsafe, and mechanisms and legal frameworks for mitigation. All really very good. Now, one part of AI that has marched forward is chips, which has fueled geopolitical friction and much else. Taiwan is the eye of the storm since it is a big and growing economy. It is where the most advanced chips are manufactured, and it is across the channel from China, whose ambitions of taking over the island are plain. So it was interesting that AWS and NVIDIA are investing in new facilities and capabilities in Taiwan. Does this mean these big companies do not see a risk of political escalation or consider their investment as too small to be held back. AI has also fueled

proliferation of chips. You can roughly divide them into five categories. The big established chip vendors, NVIDIA, AMD, Intel. Hyperscalers, AWS, Google, Azure, Meta, who have their own chips for their own use. And right behind them are smaller companies with novel approaches like Cerebrus, Grok, Samba Nova, Graphcore. You could separate out neuromorphic chips where IBM, Intel, and Rain have projects. And then you get the big consumer embedded mobile, the space that's occupied by Apple, Qualcomm, TenStorent, MediaTek, and Samsung. And the industry seems to have appetite for all of them. We hear interesting things from specialty AI chip companies that point to market traction and,

by implication, also an improving software stack. For example, Cerebrus, with its dinner plate-sized wafer-scale engine, announced last week a collaboration with Dell to integrate Cerebrus CS3 supercomputers or CS3 clusters with Dell servers that have AMD EPYC CPUs. So Shaheen, possibly this is a moment for Cerebrus and maybe the other specialty AI chip companies to move toward the mainstream and pick up greater market traction. In any case, the market looks far from consolidating. We'll see. Another big aspect of AI chips is their high energy use. Along those lines, Google announced on Friday that it plans to buy 115 megawatts of geothermal energy to power data centers in Nevada. And in the UK, there's a story that a company called

Reef Group has proposed a hydrogen fuel cell plant proposed for a 600 megawatt London data center. They note that backup power generators will run on hydro-treated vegetable oil rather than diesel. We've seen data centers no longer measured by their size, physical size, but by their power consumption, as they've gotten bigger, hotter, and more expensive to cool. Big tech companies have aggressively embraced low-carbon footprint energy sources like solar and wind, and now some of them are taking it a step further. is recent news that OpenAI, the chat GPT company, wants to buy, quote, vast quantities, end quote, of nuclear fusion energy from a startup fusion company called Halion. An interesting point being that fusion energy, according to an article in Scientific American last year, won't be available until 2050. Halion, on the other hand, says it will have a fusion power plant online by 2028. And there's one more item on AI chips.

Earlier this month, at its annual Worldwide Developer Conference, Apple unveiled a new cloud intelligence system designed for private AI processing. Apple also mentioned that they use their own silicon in the data centers as part of these new AI services they announced. This was in addition to the AI models that run on the device and the services that can access through an agreement with OpenAI. Yeah, this is very interesting because Apple's chips are very good for AI. There's a shortage of AI chips and Apple is a big customer of TSMC and obviously has massive volume. Apple already has a rack-mountable

Mac Pro, which can be configured with an Apple M2 Ultra with a 24-core CPU, 76-core GPU, 32-core neural engine, 192 gigabytes of unified memory, and 8 terabytes of SSD storage. They probably have to do something about faster interconnects, but clearly have the wherewithal to pursue that and to build the software stack. It's one of those possibilities that can change the market since it applies consumer technologies to data center uses. 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 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.