@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20250428

Episode Date: April 28, 2025

- Salishan Conferenc - HPC-AI Divergenece - FugakuNext Zettascale? - TSMC A14 Fab, Intel 18A Fab - AmKor CoWoS Packaging Arizona - Intel Earnings, AI strategy [audio mp3="https://orionx.net/wp-content/uploads/2025/04/HPCNB_20250428.mp3"][/audio] The post HPC News Bytes – 20250428 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 InsideHPC, and with me is Shaheen Khan of OrionX.net. The Salishan Conference was held last week. It's a big deal in the HPC community. It's an invitation only event that goes back to 1981 and is held at the Salishan Lodge in Oregon.

It brings together people from the national labs, academia, government, and private industry. And Shaheen, you know, the males must be bad around here or something, because I've never gotten my Salishan invitation. On the other hand, we've had the opportunity to speak with several of this year's Salishan participants at our at HPC podcast. This includes Thorsten Heffler, Dan Reed, as well as Satoshi Matsuoka, who spoke at Salishan last week on the interesting theme of AI and traditional HPC divergence. Satoshi asked whether this poses an existential crisis

for HPC or is this just a myth? Satoshi challenged the idea that AI and LLMs will undermine traditional HPC and scientific simulation, arguing that recent hardware trends, especially GPU-centric architectures, are in fact synergistic with traditional HPC. He cited work being done at the Riken Center in Japan on the Fugaku supercomputer that is actively leveraging this convergence for Fugaku Next, their next-gen supercomputer targeting zeta scale performance by 2029 or 2030. Zeta scale within four or five years? Jishin, I wish I'd been there. Well, I'm sure your

invitation is in your inbox somewhere. You got to like look more closely. Well, the subject that Satoshi covered is one of the hottest topics in HPC community today. And the discussion is a lot simpler if you believe, like I do, that AI is an HPC application and quantum computing is an HPC subroutine call. So yes, we agree with him that HPC will always be at the core of scientific discovery, as it enables increasingly sophisticated automation. Now, personally, I believe if a conference is by invitation only, then it should probably not be publicized too much and maybe even be Chatham House rules kind of a thing. But Salishan is special since there's broad interest in the material and the discussions.

The conferences line up to me is kind of like the top 500 list. Those who are there are undoubtedly big stars of the field, but many more are missing and the list is not trying to be exhaustive. Anyway, we encourage listeners to check out the Salishan website and peruse the session lineup, several of which are linked to abstracts and slide presentations and it's excellent material.

There's never been a post-Moor's law thing at TSMC. The company has been vocal about forging ahead with Mohrs law as it keeps upping the ante with smaller and smaller feature sizes and it's been reaping the rewards. It did so again this week when it announced its next generation A14 process. It's been a few years since these labels have meant the real feature size, so A14 doesn't quite mean 1.4 nanometer or 14 angstrom, but it is the next big thing. The company said the A14 process is making great progress, in fact, a bit ahead of schedule and should be in production in Taiwan in the 2028-2029 timeframe.

As we covered last week, TSMC's N2Fab in Taiwan and Intel's 18AFab in Arizona are both in the two nanometer zone and expected to go into production later this year. In addition to the actual chip, there's also a need for packaging. And that packaging is a big part of the manufacturing process for large chips that are made up of multiple chiplets. A lot of AMD chips like Ryzen 3000 and 5000 series, Epic processors, and the MI250 or

MI300 chips use chiplets. Nvidia traditionally preferred one big chip until it just got too big. High-end packaging for chiplets is called CoaS for chips on wafer on substrate and is only available in Taiwan right now. So chips that need it, like Nvidia's B200 or B200A chips, must be manufactured in Taiwan or shipped back to Taiwan for packaging. But Amcor, an old company based in the US, it was founded in 1968, one of the leaders of outsourced packaging with revenues last year of over $6 billion. And a TSMC partner with factories around the world is building a factory in Arizona for packaging and will be able to do small and then larger chiplet packaging in Arizona by 2027.

Apple will be Amcor's first customer. Nvidia mentioned working with them and another company, the Taiwanese Siliconware Precision Industry Company Limited, or SPEL, for Blackwell chip packaging and testing in Arizona. Blackwell chips use TSMC's 4NP process, which is an advanced version of their N5 5nm process,, but this one uses 4-nanometer design models and expect it to be about 10% faster. We're all watching the Intel saga. This past week saw new CEO Liput Tan's first quarterly earnings announcement, and there

was good and bad news for the company. On the upside, the company beat revenue estimates and reported at $12.5 billion in sales, roughly even with Q1 last year. But they projected a down quarter coming up, and the stock took a hit. On the strategy side, the new CEO indicated Intel will move toward a homegrown effort to compete in AI markets. In a recent message to Intel employees, Tan said the company needs to get back to its roots and empower its engineers. He also said Intel needs to revamp its culture. Quote, the feedback I have received from our customers and many of you has been consistent. We are seen as too slow, too complex and too set in our ways and we need to change.

According to the Reuters news service, the new homegrown strategy contrasts with the company's old approach before previous CEO Pat Gelsinger when Intel bought a string of chip companies like Habana Labs, Mobileye, Nirvana, and Movidius, not to mention FPGA company Altera in 2015, which we talked about last week. Of those acquisitions, only Mobileye really has built a strong position in the important autonomous driving space. But Wall Street analysts generally agree that NVIDIA's dominance in GPUs and also the work of hyperscalers to build their own AI chips taken together pose major barriers to entry for Intel in AI.

And then there's AMD GPUs. On the other hand, by indicating that Intel will develop homegrown AI chips, the company may have bought itself some time. Now, time has been in short supply for Intel, which has done well in integrated GPUs and AI PCs, but basically sat out the server GPUs that fuel most of AI today. Intel has had all the right ideas,

including building rack-scale systems and even internet data centers back in the 90s, well before grids and clouds emerged. But it let others get ahead and needs to catch the next wave. We note too that Nvidia and AMD report earnings in a couple of weeks. Let's see if Nvidia can blow the roof off expectations once again. All right, that's it for this episode. Thank you all for being with us. HPC News Bytes 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.