@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20231225

Episode Date: December 25, 2023

- EU Exascale, HLRS, MareNostrum Inauguration - Neuromorphic AI gets active - ASML's 1st High-NA EUV system goes to Intel - Argonne's Bimetallic All-optical Switch [audio mp3="https://orionx.net/wp-content/uploads/2023/12/HPCNB_20231225.mp3"][/audio] The post HPC News Bytes – 20231225 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. Europe is moving into the exascale era. They're coming in several years behind China and the U.S., but there are some advantages to

their timing. For one thing, their projected costs are significantly less than the $600 million price tags per system, more or less, paid by the U.S. for its first three exascale systems. So the news is HPE and the University of Stuttgart have announced they will build two new supercomputers at the HLRS Supercomputing Center. The first system, a sub-exascale system called Hunter, will begin operation in 2025, followed in 2027 with the installation of the exascale class Herder. The total combined cost for both systems is 115 million euros. Also, the Barcelona Supercomputing Center announced the inauguration

of the Marenostrum supercomputer. The Eviden Bolsaquana system is listed at number eight on the top 500 list at 314 petaflops. Deep South, a system described as the world's first brain scale supercomputer is planned for operations four months from now, in April 2024. It will be located at the International Center for Neuromorphic Systems, ICNS, at Western Sydney University, in collaboration with University of Sydney, University of Melbourne, and University of Aachen in Germany. Now, neuromorphic chips are motivated by the human brain and its low energy use on the order of 20 to 30 watts as compared to 500 watts and increasing for current AI chips. They emulate large networks of biological neurons and their synaptic operations. Neurons, of course, are brain

cells that process and transmit signals through synapses that act as bridges between neurons so the metric for these chips is synaptic operations per second other notable neuromorphic chips are IBM's True North that has been around for a very long time Manchester University's Spinnaker chip and system also has been around for a long time a startup called Rain in Silicon Valley and various research in academic world. The field is not viewed as ready for prime time, but it has got more active recently in pursuit of low energy solutions. Yeah, really, this is an amazing announcement in part for the promised system performance,

228 trillion operations per second. That's better than most humans on a good day, isn't it? There's not a lot of specifics about the system that were released, though the researchers say Deep South will leverage FPGAs, enabling the addition of new neuron models, connectivity schemes and learning rules, which they say overcome limitations in other neuromorphic computing systems with custom-designed hardware. In fact, Intel has a nice use case on Western Sydney University using Intel FPGAs in a system, and Dell mentioned similar research to simulate neurons. ASML, the Dutch company, announced it shipped its first high NA EUV system to Intel this week. This is a twin-scan EXE5000 extreme ultraviolet scanner. It's a pilot system at this stage.

It's ASML's first high NA system. And it's interesting to note that Intel placed the order for the system in 2018. Intel will use it to experiment with high NA EUV before deploying it for volume manufacturing in 2025. Yeah, I think Intel receiving the first high NA system is very interesting. Since it indicates A, Intel was taking steps to catch up earlier than commonly believed, and B, because Intel is receiving the first one, while TSMC is viewed as the closer partner with ASML with deep engineering ties. Of course, the historical connection between ASML and Intel are pretty strong. Back in 1999,

ASML was part of a consortium with Intel and others to license and use fundamental research from the USDOE. And in 2012, Intel invested $4.1 billion in ASML for a 15% equity stake. That stake is now down to about 3%. TSMC, by the way, also had an equity stake, but sold its entire stake in ASML in 2015. Optical technologies are used in telecommunications and fiber connectivity, but also making advances in silicon photonics. They promise lower energy costs and higher bandwidth, so it was encouraging to see news that scientists at Argonne National Lab have created a bimetallic optical switch that can both transmit data and store it. In that announcement, an Argonne scientist was quoted saying,

when you use optical components instead of electronic circuits, there are no resistive capacitive delays, which means that in theory, you could operate these chips a thousand times faster than conventional computer chips. So really exciting stuff. All right, that's it for this episode. Thanks so much for being with us and have a great holiday week.