@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20241209

Episode Date: December 9, 2024

- Intel CEO Pat Gelsinger retires - Marvell-AWS supplier alliance - Chinese quantum computer Tianyan-504 [audio mp3="https://orionx.net/wp-content/uploads/2024/12/HPCNB_20241209.mp3"][/audio] The post HPC News Bytes – 20241209 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. I'm Doug Black. Welcome to HPC News Bites. I'm with Shaheen Khan of OrionX.net. The big news of the week is Pat Gelsinger leaving Intel, and it's kicked off a tsunami of commentary and speculation about why he left the company and what comes next. Much of the talk has focused on Gelsinger taking Intel into the foundry business, meaning they would not only produce chips for themselves, but also for others.

No doubt it was a highly expensive and time-consuming move that contributed to big layoffs at Intel, but even so, on the chip design side of the equation, Intel remains behind AMD on server CPU technology and far behind NVIDIA on GPUs. Gelsinger promised Intel's return to technical leadership, but it hasn't happened yet. It's undeniable that Intel's problems go back well before Gelsinger returned to the company in early 2021, including the decision 12 to 15 years ago to leave the emerging GPU market to NVIDIA. This catapulted them into AI dominance and the revenue stratosphere. As for AMD, their renaissance under Lisa Su's leadership began before Gelsinger took the reins at Intel. AMD beating Intel on CPU price performance has resulted in that company overtaking Intel on data center CPU revenues. Where all this leaves Intel and who will take the reins is all yet to be seen. Shaheen, what thoughts do you have on the Gilsinger news? Yeah, many thoughts. At the high level, I think companies with an explicit focus on HPC have been on the winning side. We

are living in a digital world, and that means lots and lots of data, and manipulating all that data is HPC one way or another. NVIDIA focused on matrix multiplies with their graphic chips, and they've kept at it, and that turned out to be just right. It led to simulation and eventually to AI. AMD has talked about high-performance computing explicitly as a strategy, and they're doing well too. Intel's focus on HPC always appeared as a sideshow, subordinate to some other umbrella strategy. In its day, IBM's Power CPU appeared to miss the deep learning revolution and seemed to intentionally relegate accelerator co-processing to NVIDIA for simulation work, not realizing AI was going to be all about accelerators. Back to Intel. Few people are

privy to whatever boardroom drama that took place, and depending on how you see it and your assumptions, you'll interpret it differently. My perspective, as you know, is that Intel's turnaround was making excellent progress and needed to be left alone on national security grounds, that it was going to take way longer than three years. Some say Gelsinger should have, as a condition of taking the job, installed a new board that understood and was aligned with his strategy and the time it would take. There's been talk of splitting the company now into two or three pieces, like the spin-out of their FPGA business, but that's going to be very tricky. Regardless, I hope the nation's progress on semiconductor manufacturing continues

somehow despite this disruption. It needs to. Sticking with chip news, the Wall Street Journal ran a front page story yesterday on the rapid growth and rising share price of Marvell. In fact, it's another chip company that has surpassed Intel's valuation. And this is due in part to its long-term deal to develop custom AI chips for AWS. And AWS, of course, made big news last week at their reInvent conference, including news around its Tranium AI chips. The hyperscalers are looking for alternative suppliers from NVIDIA for custom AI chips, a market segment one analyst firm says could reach $30 to $50 billion by 2030. The journal article noted that Marvell's CEO, Matt Murphy, who has been in that

spot since 2016, took the unusual step of taking a moment out of his recent earnings announcement commentary to deny rumors that he may be headed for the top spot at Intel. As hyperscalers morph into system vendors, well, it seems like everyone is morphing into system vendors, they need a full supply chain. So it makes sense to create these alliances and reserve capacity because everything is in short supply too. Marvell, Microchip, and Mediatek are examples of semiconductor company names we will hear more often. Marvell has been around for 30 years, based in Santa Clara, California, lots of acquisitions along the way, revenues of around $5.5 billion in their fiscal 2024, and about 6,500 employees and operations in many countries. They make all kinds of chips. Their first big customer way back was Seagate, so you continue

to see them in the path between storage and networking and the CPU. This includes up to electronics now, and also the Tranium chip for Amazon. Marvell's stock price has skyrocketed as the market for custom chips explodes. Under a five-year agreement with AWS, Marvell provides a broad range of data center semiconductors, including Tranium and its variants. The Journal also reported rumors of Marvell building an AI chip for another unnamed big tech customer that analysts believe to be Microsoft. Let's turn to China. We all know from the top 500 list that in recent years, China has been keeping relatively quiet about its advances in supercomputing and perhaps other technologies.

So it's notable when there is any news, either embedded in scientific journals or an explicit news report. But often the news lacks details and third-party corroboration. So there was news in Chinese media this week that touted what they called, quote, a new domestic record, end quote, with the launch of their Tianyan 504 superconducting quantum computer. It's a 504 qubit Xiaohong chip and compares well with superconducting quantum systems from IBM, Google, IQM, etc. The action in quantum computing has been on error correction to get more out of existing systems, which is necessary for scale, and also on modularity, the idea of connecting multiple quantum chips together to build a bigger one. The news is not surprising, but I wish it included a

lot more detail. The Tianyan 504 surpasses the 500 qubit chip threshold and also promises to match the performance of international quantum computing platforms like IBM in terms of qubit lifetime, readout fidelity, and other key aspects. This quantum computer will be integrated into the Tianyan Quantum Computing Cloud platform, making it accessible to users worldwide. So we should get a pretty good idea pretty soon what the Chinese have accomplished. Okay, that's it for this episode. Thanks so much for being with us. HPC Newsbytes is a production of OrionX in association with InsideHPC. Okay, that's it for this episode. Thanks so much for being with us.