@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20231127

Episode Date: November 27, 2023

- NVIDIA Ethernet push with Dell, HPE, Lenovo - Research and Engineering Studio on AWS - Latest on Chinese exascale - HPC-Quantum integration: Riken's Fugaku + NTT Ei and Simulated systems [audio mp3="https://orionx.net/wp-content/uploads/2023/11/HPCNB_20231127.mp3"][/audio] The post HPC News Bytes – 20231127 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. Hi, Shaheen. It was a quiet week in the world of HPC AI, to paraphrase Garrison Keillor, week after SC in the week of Thanksgiving, other than, of course, the incredible drama surrounding open AI, which captured the attention of the week of Thanksgiving. Other than, of course, the incredible drama surrounding open AI, which captured the attention of the world's media. But in the way of product news, NVIDIA announced Ethernet for AI on Dell, HPE, and Lenovo platforms, which will integrate NVIDIA Spectrum X Ethernet networking. NVIDIA said it is designed to speed up generative AI

workloads, achieving 1.6 times higher networking performance for AI communications. Moving data is the most expensive part of computing just when all applications are going big data. Now, we used to say what I just said about 10 years ago, and folks like John Gustafson have been saying that since early 1980s that I could locate. It's been a looming challenge. GPUs themselves are a way to get around it. And mixed precision is another way. And optimizing the end-to-end stack is an ongoing effort,

most recently exemplified by the Ultra Ethernet Alliance. Optimizing specifically for AI, however, is something that you should take with a grain of salt since the whole area is evolving and data movement and access patterns have not gelled yet. But fast is good and optimizing it in various ways is also good. Another piece of it's deja vu all over again is news of an HPC portal over at AWS to simplify administration and running of applications in the cloud. It's called Research and Engineering Studio, RES, and is able to create a Linux or Windows virtual desktop

with applications ready to go and collaboration tools like shared data. Yeah, I talked with AWS at the conference. They certainly emphasize the administrative advantages of the offering. Managers can set permissions and budgets. The real-time budget oversight seems to be a major selling point. The avoidance of surprise cost overruns. There was also news from China at SC. Sunway showed its SW20610 Pro processor, according to an article in the Chips and Cheese publication, which dates back to late 2020 to early 2021.

And it's an improved version of the chip featured in the Sunway Tiahu Lite. I know you have some thoughts on this, Jaheim. I would, yes. So the existing system at number 11, which showed up in 2016, has over 10.5 million cores, no accelerators, and scored 93 petaflops, which was 74% of its peak speed. This system looks like a linear progression, more core groups, more nodes, and apparently more focus on mixed precision also. It comes in at over 41 million cores, so at about the same frequency and efficiency, it should do about 350 to 450 petaflops for HPL. One should also look at the harder HPCG conjugate gradient benchmark, which is a kind of lower bound on

performance. We didn't cover it with our top 500 episodes, we should. If you're getting more than about 1% to 2% of your HPL performance on HPCG, you're doing well. So it's that much harder and a little bit more indicative of what you can get for average workloads. Another piece of news was the Rican Center in Japan announced that they're taking the next step in their quantum science, creating what they aim to be, quote, world's largest and most capable quantum HPC platform, coupling Fugaku with the world's largest quantum machines, real and simulated. So this one is very specifically focused on integration, and it's consistent with what

other supercomputer centers are doing, most notably ULIC in Germany, which has a D-Wave and Eviden and Pascal quantum capabilities, the DOE centers in the US, Barcelona, CEA in France, Chineca in Italy, Poznan in Poland, and IT for Innovation in Czechia all have quantum computers also. So as more and more supercomputer centers integrate quantum computers,

the time is coming to dust off the label integrated heterogeneous supercomputing that came out of floating point systems in early 1990s. But the trend is coming to dust off the label integrated heterogeneous supercomputing that came out of floating point systems in early 1990s, but the trend is coming back. Yes, I know that's a favorite of yours, Shaheen. RIKEN says it will focus on integrating quantum computing capabilities on the Fugaku ARM-based supercomputer at the RIKEN Center. Part of this effort will include development of software to link quantum and supercomputing. In addition, RIKEN is conducting R&D on collaboration between the first superconducting quantum computer in Japan called EI and Fugaku.

All right, that's it for this episode. Thanks for being with us. HPC Newsbytes 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.