@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20251222

Episode Date: December 22, 2025

- High Frequency Trading race to save nanoseconds - AI federated learning at DOE labs - Perspective from "AI nobility" who are also "AI doomers" [audio mp3="https://orionx.net/wp-content/uploads/2025/12/HPCNB_20251222.mp3"][/audio] The post HPC News Bytes – 20251222 appeared first on OrionX.net.

Welcome to HPC Newsbytes, a weekly show about important news in the world of supercomputing, AI, and other advanced technologies. Hi, everyone. Welcome to HPC Newsbytes. I'm Doug Black of Inside HPC, and with me is Shaheen Khan of OrionX.net. A classic enterprise application of HPC is high-frequency trading, in which investors use supercomputers to achieve stock trading speed advantages measured in milliseconds. But now, traders on a European futures exchange are shaving that time frame down to a million times faster than a millisecond. A story last week in the Wall Street Journal looks at Germany's Eurex exchange and a dispute in which a group of traders have used garbled data to save

3.2 billionths of a second. As the journal explains it, the controversy is about an arcane technical maneuver in which high-speed traders bombard Eurex with useless data. The idea is to keep their connections to the exchange warm so they can react fractionally faster to market-moving information. The controversy became public when the Paris-based trading for Mosaic investigated why that firm's formerly thriving trading business suddenly fell by 90% in 2022. The journal quotes the founder of Mosaic saying, an arms race is okay, but you must use legal weapons. Well, a program trading, algorithmic trading, straight through processing, and finally high-frequency trading, HFT, are part of the spectrum of the discipline to execute financial

trades faster and faster. You have financial trading companies at one end and market data at the other. What connects them is the network and the level of optimization that goes into that is just unreal and has been for decades. There's a very good paper from earlier this year that covers this very well. It's in the International Journal of Computer Science and Network Security and it's titled, quote, the impact of high-speed networks on HFT performance. It is acceptable. It is accessible online. In this case, the optimization goes all the way into network packets that tell the market a trade request is coming, thus keeping it warm, only to decide whether or not to actually issue the transaction depending on market data. Like you said, it's down to milliseconds or less.

As the Department of Energy's Genesis mission rounds into form, there have been announcements about participating organizations. The number is approaching 30, and we're seeing some of the strategies that may be used to construct the planned cooperative array of supercomputing resources at various locations around the country in support of their AI for science mission. One of them is a federated learning AI model project led by Sandia National Laboratories that the lab calls a significant milestone in advancing AI for national security. Over the past year, Sandia, Los Alamos, and Lawrence Livermore Labs, known as the Tri-Labrador, have been building a federated AI model as a pilot project, and they now have a prototype.

A federated AI model trains a model that is shared across devices in multiple locations without moving the raw private data with the goal of creating a smarter collective global model. The project used an open-source federated learning software called NV flare, and as you can tell from the name, it's contributed by NVIDIA. The distributed learning proceeds in parallel at each lab and is broken up into phases or epochs. And after each one, the labs share the updated weights, but not the data itself. And the weights are averaged together to form a single model for the next epoch of training to begin. Continuous and federated training is one of the next stages of AI and critical for achieving high-quality learning and therefore inference.

ability to use classified datasets for training without sharing them is itself a critical requirement for government and commercial applications. Tech media mostly concerns itself with how AI technology is evolving and advancing with less focus on the possible negative ramifications of AI as it approaches artificial general intelligence and superintelligence levels of performance. Certainly and naturally, we don't hear concerns from AI vendors whose mission, of course, is full speed ahead. But there is a serious group of technology thinkers, analysts, and academics, deeply troubled by the possibility of dystopic AI. This group has been labeled the Doomers. And MIT Technology Review has published a lengthy article on the Doomers perspective and how they've adjusted their

outlook as AI progress has apparently stumbled a bit of late. We're referring to the somewhat disappointing reception given to GPT-5 from OpenAI, whose CEO had said in 2024 that AGI may arrive this year and superintelligence by 2030. At the heart of the Doom Review is a call for technologies, measures, and policies for greater AI safety before the industry developed systems it can't control. From their perspective, AI's slowed progress merely means the world has more time, a bit more time, to come to grips with AI. I'll quote, a characteristic statement from the Dumer Group, Joshua Benjio, winner of the Turing Award and chair of the International AI Safety Report said, quote, the overall landscape for AI governance

and safety is not good. There's a strong force pushing against regulation. It's like climate change. We can put our head in the sand and hope it's going to be fine, but it doesn't really deal with the issue. He later added, like many people, I had been blinding myself to the potential risks to some extent. You're excited about your work and you want to see the good side of it. That makes us a little bit biased in not really paying attention to the bad things that could happen. Even a small chance like 1% or 0.1% of creating an accident where billions of people die is not acceptable. It's the global question of our times and it's an important article. because it highlights how serious the industry's views are.

AI is already impacting businesses and societies. Some people believe that policymakers should make sure short-term impacts of AI are covered before focusing on medium and long-term issues. Short-term impacts of AI, like deepfakes, clever cyber attacks, job elimination, or an AI bubble, get a good amount of attention these days. But all aspects of AI policy are important,

and this article discusses the long-term, impact. AI is not done yet, of course, and progress needs to continue for both business and social benefits. It's not easy to manage risk in the face of something that is unknown and transformative. It leads to the evaluation of how any unexpected negative consequences are managed versus how unexpected positives are enjoyed. For example, Nobel Prize winner Jeffrey Hinton said that his focus has been on the longer term threat, quote, when AI gets gets more intelligent than us, can we really expect that humans will remain in control or even relevant, question mark, end quote. There's also a lack of an accepted framework with probabilities

of something happening or not and timeframes needed for mitigating a threat. You see Berkeley Professor Stuart Russell says it, quote, isn't about imminence. If someone said there's a four-mile diameter asteroid that's going to hit the Earth in the year 267, we wouldn't say, remind me in 2066 and we'll think about it, end quote. Now, it has been said that logic is always on the side of the naysayer, but on the positive side, the article says, quote, most people I spoke with say their timelines to dangerous systems have actually lengthened slightly in the last year, an important change given how quickly the policy and technical landscapes can shift. Let's hope for that. 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.