@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20250512

Episode Date: May 12, 2025

- SMIC, China's Semiconductor Industry - Cisco in quantum networking, quantum scale-out - HPSF.io open source HPC stack - The journal Nature sees US science brain drain as EU allocates €500m to attract international scientists [audio mp3="https://orionx.net/wp-content/uploads/2025/05/HPCNB_20250512.mp3"][/audio] The post HPC News Bytes – 20250512 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 of InsideHPC, and with me is Shaheen Khan of OrionX.net. The Wall Street Journal ran an upbeat article last week about the Chinese chip manufacturer SMIC. It talks about a profit jump and quote, a sharp rise in quarterly net profit. But Shaheen, I came away with a less positive impression. For one, the article doesn't mention

AI or data center. SMIC is focused on consumer electronics and they're still at 7 nanometers. In terms of scale, the first quarter revenues were $2.25 billion, which amounts to about a tenth of TSMC's revenues on an annualized basis. Yet the Journal stated that SMIC plays a key role in Chinese push for chip self-sufficiency. If this is the best the world's second largest economy has at a time when AI is the primary arena of competition, I have my doubts. But what say you? Well, from 2015 to 2020, China's semiconductor industry tripled in size, going from $13 billion to $40 billion. That was before trade tension started, but the consensus is

they have been building strength and marching forward. China lists thousands of companies in the semiconductor space, maybe even tens of thousands, according to some reports, and a lot of public investment into the space. SMIC is the leading pure play foundry. High Silicon is a large and growing subsidiary of Huawei in Fabless design. Jset, J-C-E-T, is in assembly and test. YMTC in 3D NAND flash and CXMT in DRAM. According to their Q1 2025 report, SMIC grew by 28% year over year, but only 2% quarter over quarter, and they projected revenue and gross margin decline for the second quarter.

Most of SMIC's revenue comes from within China, 84% in the last quarter. Capital investment, they said in 2025, would remain the same as last year, about $7 billion. Automotive and industrial segment grew by 71% year over year, but that was only 10% of total revenue. Then there is smartphones, consumer electronics, some Bitcoin mining chips, and the AI chip

that they are manufacturing for Huawei. That AI chip, as we've covered here, is the Ascent 910C, 910C, which is about a third of the performance of the Nvidia B200, but Huawei has stuffed 384 of them in their Cloud Matrix 384 system. That's five times as many as the 72 GPUs in Nvidia's GB200 NVL72 system. That also means the Huawei system needs four times the power compared to NVL72. And it's generally harder to get performance out of 394 chips compared to 72, but it puts them in the game. And then you have advances, like we saw with DeepSeq,

which points to the possibility of getting more out of slower systems than before. So the race is very much on. SMIC is using old and not so old fab technologies for different industries. They have been building in seven nanometers and had a development project, presumably using internally developed and indigenous technologies to get to five nanometers. The five nanometer project was targeted for completion by the end of last year. So we may see five nanometer chips this year or next. By comparison, TSMC and Intel

are expected to go into production with 2nm lithography this year. So make it several years behind. Cisco came out with an interesting announcement last week that sounds like a Beowulf cluster strategy applied to quantum computing. The company said its new quantum networking technology will connect smaller quantum nodes and make quote quantum computing practical years ahead of current timelines. The head of Cisco's internal incubation group called

Outshift said the key to practical quantum computing is connecting nodes using network infrastructures to create distributed systems. The future of quantum does not lie in a single monolithic quantum computer, they said. Of course, current quantum processors are limited to hundreds of qubits, while practical quantum applications require millions. But Cisco says its quantum networking technologies will allow quantum scaling and claims that quantum will be a practical reality, not in decades,

but five to 10 years. It's an interesting concept, Shaheen, I'd say. I just wonder if HPC class clusters is a transferable idea for quantum. It very much is. Even on their own turf, i.e. for problems that can use a quantum computer, quantum computers need to scale if they want to beat GPUs. So you either make a big chip with thousands of qubits on it or you string together existing QPUs for quantum processing units. So it's the same scale out versus scale up ideas but with different dynamics. And for quantum scale out, quantum interconnect is emerging fast and we see companies develop it, acquire it, or position themselves as agnostic. Cisco formed the Quantum Networking Lab in Santa Monica, close to the Los Angeles

airport, and also talked about a prototype chip that generates entangled photon pairs. That chip was done together with UC Santa Barbara, which is also where Google's main quantum research is, complete with a FAP facility and a data center. As you recall, quantum measurements and quantum communications are already important markets. So quantum networking has a lot of applications it could pursue. Quantum key distribution, quantum secure networking,

tamper-proof location authentication, more accurate synchronization using quantum sensing, range extension using quantum repeaters, etc. There are a lot of projects in this area that have little to do with a programmable quantum computer, and quantum scale-out is farther out. Now there are probably over a hundred umbrella organizations around the world that drive open-source software. Think of them as legal vehicles, each structured to match the focus, governance, financial model, development process, collaboration model, etc. to the kind of software they aim to

foster. Notable foundations include Apache, Cloud Native Computing, Linux, Mozilla, Eclipse, OpenBSD, OpenSSF, Wikimedia, etc. The list is really pretty long. Back at SC23, the Linux Foundation announced it was going to form the non-profit high-performance software foundation HPSF to focus on open source HPC software. HPSF was formed at ISC24 in May of 2024 with 16 members and has added seven since. Premier members include Lawrence Livermore National Lab, AWS, HPE, and Sandia. But the full list includes AMD, ARM, Intel, NVIDIA, and many others. Yeah, this is important for the traditional HPC community as well as enterprise IT. AI is steering the hardware and software

architecture of data centers everywhere, while the HPC software stack is blending with the traditional commercial enterprise software. This type of effort can facilitate the infusion of AI into enterprise applications. HBSF had a big conference last week and at this point covers a pretty nice set of software layers, including SPAC, which is a software package manager that was driven by Lawrence Livermore National Lab, CoCoS, a hardware agnostic development model for C++ applications, HPC Toolkit for performance analysis, and then Abtainer for HPC-inspired containers, what used to be called

Singularity, E4S, which stands for Extreme Scale Scientific Software Stack for scientific libraries, OpenShammy for management, and Whiskers for scientific visualization. HPSF.io is the website. Go check it out. We'll end with news of a report in the journal Nature titled, quote, exclusive, a nature analysis signals the beginning of a US science brain Nature titled, quote, exclusive in nature analysis signals

the beginning of a US science brain drain, end quote, saying US scientists are increasingly considering jobs in other countries and international applications for US science jobs are dropping. Separately, the European Union said it's allocated 500 million euros in 2025 to 2027 to attract researchers from outside of Europe. 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 InsideHPC. Shaheen Khan and Doug Black host the show. Every episode is featured on insidehpccom and posted on OrionX.net. Thank you for listening.