@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20240909

Episode Date: September 9, 2024

- Copyright law, Fair Use doctrine, and AI's appetite for data - TSMC's CoWoS capacity - Intel's 18A fab - DOE's $23 million RFP re future of HPC [audio mp3="https://orionx.net/wp-content/uploads/2024/09/HPCNB_20240909.mp3"][/audio] The post HPC News Bytes – 20240909 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. Hey, everyone. Welcome to HPC News Bites. I'm Doug Black of Inside HPC, and with me is Shaheen Khan of OrionX.net. Shaheen, we'll start with a story that could have important implications for generative AI, large language models, and their reliance on enormous data sets pulled from the internet to run gen AI workloads. The U.S. Court of Appeals has ruled in favor of a group of booksellers,

upholding an earlier ruling that the non-profit internet archive organization's book digitization projects violated copyright law. There's a saying in copyright circles that content wants to be free, but really that usually applies more for consumers of content than producers of it. They basically said wholesale digitization of books is not fair use. If you interpret that as large-scale ingestion of data in general, then it starts being relevant to AI companies. The words fair use refer to a legal principle, a doctrine, that says in certain circumstances and in limited ways, copyrighted material can be used without license or payment. It is meant to reduce friction, and it is often mutually beneficial. But this is a thorny area, since existing laws never imagined that data can be

used for the kind of impact, profits, and also valuations that AI companies are pursuing. So it's understandable that data creators would want a share of that and want to clarify or change the laws. But the way AI models use data is complex. It includes lots of mathematical transformations and processing. I see the big idea here as the data supply chain, and we need clear laws that govern data rights as it flows through the data supply chain and becomes more valuable. Advanced chips are getting larger in every dimension. They call it 2.5D when it is horizontal and 3D when it is vertical stacking. That has led to chip-on-wafer-on-substrate COBOS, C-O-W-O-S, a system-in-package, S-I-P, technology that brings chips or chiplets closer together, that improves flexibility,

speed, and power consumption. It was first introduced in 2012 and has advanced steadily. It is used for high bandwidth memory, for example, and is increasingly interesting for larger chips. In short, it's hot, it's in demand, and surprise, surprise, there's a shortage of capacity. TSMC is doing something about that, too. It was reported this week that TSMC is seeing enough demand to grow their COAS production capacity by more than 50% every year from 2022 to at least 2026. That's a nice compound annual growth rate. This is according to a report this week citing TSMC's Vice President of Operations Advanced Packaging Technology. Amid disappointing news from Intel that has dragged Intel shares below $20, the company released some good news this week. Reuters ran a story in which Intel said its foundry business will realize, quote, meaningful revenue

in 2027. This could be built on current packaging business where Intel has advantages and one would hope also from actual fabrication. Intel also said that since releasing its 18A process design kit in July, they are, quote, encouraged by what we're seeing in the fab. It's powered on and booting on operating systems and yielding well. And we remain on track for launch in 2025. The chip side and the manufacturing sides of Intel are pursuing what they need to do to be successful. That means the chip side has started using TSMC when it makes sense, and the fab side is pursuing other customers for the capacity they are building. That all is very good. The big good news here

really is Intel's impressive progress in fabrication technology. This is what many people said was impossible, and some are still saying that. It's hard to do. Their roadmap was Intel 3, then Intel 20A, 18A, and 14A, some of it pursued in parallel. Intel said they would skip 20A and go straight to 18A in 2025, which indicates confidence in how 18A is going. There was news that 18A had glitches with a Broadcom chip, but that is not uncommon at this stage. By the way, Intel 18A will be very competitive with TSMC. Now, if they skip 20A, then their own chips relying on that process will probably end up being manufactured by TSMC, and that's not a bad thing either. The Department of Energy's Office of Science has issued a $23 million RFP to, as they say, define the future of HPC, that is, to advance

technologies for future supercomputers. DOE said the New Frontiers program will form partnerships with companies for research on critical technologies used in post-exascale systems for 2029 and beyond. Well, it's called high-performance computing for a reason. It's a permanent push to open up bottlenecks that impede performance. The OE specifically mentioned energy efficiency, advanced memory, interconnects, and programmability as key focus areas. So for the visionaries and technologists amongst you, the RFP deadline is October 21st. You got time to respond. All right.

That's it for this episode. Thank you all for being with us. HPC News Bites 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.