@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20241223

Episode Date: December 23, 2024

- Long-Thinking AI - UALink + optical - Arm-Qualcomm lawsuit [audio mp3="https://orionx.net/wp-content/uploads/2024/12/HPCNB_20241223.mp3"][/audio] The post HPC News Bytes – 20241223 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 Inside HPC, and with me is Shaheen Khan of OrionX.net. Long thinking could be the next big thing in AI. It's a term used by Jensen Wong during a recent NVIDIA earnings call. And according to an article in the Wall Street Journal, it, quote,

has the potential to reduce or eliminate the errors that frequently pepper the instant responses generated by large language models like ChatGPT. Long-thinking AI models, quote, think over, unquote, the results they generate, and they have the intelligence to give updates on their progress and request feedback as they do their work. The journal noted that OpenAI's long-thinking technology was seen in September with the launch of their O-series models, which take longer as they reason through complex tasks and take on harder problems than previous models in science, coding, and math. We need several more milestones in the world of AI as progress continues, and long thinking expands on two types of human thought. What psychologist Daniel Kahneman, who passed away earlier this year and had won a Nobel Memorial Prize in

economics, called System 1 and System 2. The analogy used to describe this is a professional camera that has an automatic setting for quick shots, but it also has various manual controls for expert photographers to go beyond the quick take and optimize for particular situations or effects. Long thinking is only one approach under development to enhance AI outcomes and decisions. Another is having multiple AIs or AI agents work as a team, as we discussed in a previous episode, or using specialized and generalized models together on the same problem. A market is created every time there is a successful standard in

networking, and there's a wide spectrum for such standards, given requirements for distance, bandwidth, latency, coherency, whether it's wired or wireless, etc. At the top end, we have chiplets, then motherboards, racks, rows, across the data center, campus-wide, all the way to the whole internet. That means standards like UCIE, PCIE, down to Ethernet, and everything from fiber optic cables to undersea cables, cellular 5G and LTE, and then Wi-Fi, Bluetooth, ZigBee, NFC, and several others. The market is ignited when companies join the standard and signal that they will be investing in it and products are coming. The UCIE standard, UCIE stands for Universal Chiplet Interconnect Express, for example, is supported by all the companies that are doing chiplets. AMD, ARM, Intel, Samsung, TSMC, NVIDIA, and many others. Right below that, we have UA-Link, driven by the Ultra Accelerator

Link Consortium, and focusing on high-speed, low-latency, high-coherency interconnects for accelerators in a scale-up mode. NVIDIA was already down this path a few years ago with its own NVLink, so this standard needs to catch up to NVLink before NVIDIA might adopt it. Yes, the standard has been adding companies that have significant and focused development on optical technologies. LightMatter is the newest member in addition to Light Intelligence, AlphaWave Semi, Astera Labs, and Credo. LightMatter said it will apply its 3D-stacked photonic engine for accelerator-to-accelerator communications, enabling high-performance and AI clusters. As we see Silicon Photonics interconnects edge toward viability, Light Matter's work with the UA-Link may be a step in the direction of commercial adoption of what could be a revolutionary interconnect technology. In the blockbuster licensing case brought by ARM against

Qualcomm, a jury last Friday unanimously decided on two of three counts in Qualcomm's favor, but it deadlocked on whether Nuvia, acquired by Qualcomm in 2021, breached its own agreement with ARM. Now ARM plans to demand a retrial on that part of its lawsuit. Qualcomm won its legal fight with ARM over charges that it violated licensing agreements related to its Snapdragon X processors, which feature Orion cores originally developed by Nuvia for data center processors under a different license agreement with ARM. Qualcomm has been a long-term licensee of ARM and does not lack in legal expertise, so the common presumption is that its license covers what it needs to. Also, Qualcomm, via its subsidiary Qualcomm Technologies Inc., acquired a company called Nuvia in 2021 for $1.4

billion. Nuvia had its own license with ARM. Legal contracts are like Turing complete code, and they need to be airtight. Sometimes it's hard to verify that they cover exactly what they need to cover, no more, no less. So when they are silent or ambiguous about an issue, they are open to interpretation, and you're liable to end up in court to try to decipher it. According to Tom's hardware, Arm was, quote, unhappy that Nuvia's technology would be used for client devices. Also, Arm claimed Nuvia's licensing term did not transfer automatically with the acquisition by Qualcomm and demanded the terms be renegotiated, end quote. While the two counts unrelated to Nuvia were judged in favor of Qualcomm,

the jury didn't reach agreement on whether Nuvia breached the terms of its licensing agreement with ARM. This is the basis of its planned request for trial. All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes is a production of OrionX in association with InsideHPC. Shaheen Khan and Doug Black host the show. Every episode is featured on InsideHPC.com and posted on OrionX.net. Thank you for listening.