HPC News Bytes – 20261019

SPEAKER_1 (0:05)

Welcome to HPC News Bytes, a weekly show about important news in the world of supercomputing, AI, quantum computing, and other advanced technologies.

Doug Black (0:14)

Hi, everyone. Welcome to HPC News Bytes. I'm Doug Black, and with me is Shahin Khan.

The Department of Energy's Genesis mission, the massive project involving several national labs, supporting AI for science and other scientific research, is moving from concept toward execution. DOE announced 12 new phase two projects, totaling $159 million, along with six additional phase one projects covering a broad range of scientific challenges. CommonWealth Fusion System will develop an AI-enabled digital twin for its Spark SPARC Fusion System. Fermilab will work on designing microchips for extreme environments. Argonne National Lab will use AI to modernize and optimize scientific software, while other projects address quantum error correction, particle accelerators, material science, and rare earth extraction. DOE also announced a Graduate Fellowship Program to develop scientists with expertise in both AI and their scientific disciplines.

And on the software side, DOE has released Genesis Agent Skills, a collection of reusable tools and guidance intended to help researchers develop AI-enabled scientific workflows.

Shahin Khan (1:37)

Genesis is one component of what we here have been calling a national discovery infrastructure.

Add to that the National Artificial Intelligence Research Resource, or NAIRR, initiatives at the National Science Foundation, National Laboratories, Universities, and then scientific instruments, and the computing and data resources that support them. And the opportunity is to make these resources work together more effectively. Genesis will push toward that while also bringing in greater participation from private industry. It's a journey though. DOE's language about superintelligence and the new golden age of science sets out an ambitious vision as they march from funding projects to building the systems and then on to integrating infrastructure and ultimately demonstrating accelerated scientific discovery.

Doug Black (2:27)

Researchers from Microsoft, MIT, and collaborating institutions have published a new method for improving weather forecasts in the period between two and six weeks into the future. The paper, published in Nature, introduces probabilistic bias correction, or PBC.

It uses machine learning to identify systematic errors in historical forecasts and correct the probability distributions produced by existing weather models. Applied to the European Center for Medium Range Weather Forecasts Systems, PBC doubled the modest sub-seasonal forecasting skill of its AI model. It also improved the performance of the center's conventional forecasting model across 98% of precipitation targets and 92% of temperate targets. The researchers tested their approach in a real-time international forecasting competition last year, where it outperformed competing systems across the evaluated weather variables and forecast horizons. The software has also been available as open source.

Shahin Khan (3:36)

Yeah, weather forecasting is one of the classic applications of supercomputing, of course. For decades, progress has come from better physics, richer data, finer simulation grids, and of course, faster supercomputers. The standard progression of forecasting windows is short-term, one to three days, medium-range, three to ten days, sub-seasonal, which is this research, two to six weeks, and seasonal beyond that. They are all hard problems, but they get much harder, and predictability generally decreases as you extend the range. More complex phenomena, like El Nino and hurricanes and such, add other complexity and often require their own specialized models.

And the community has been looking at using AI as a tool for several years now, well before LLMs came on the scene. What's new here is that the researchers built a relatively lightweight AI-enabled correction layer on top of existing models. Like you said, they looked at decades of data comparing what the simulations predicted and what the actual weather was, and used AI to identify what might be described as systematic biases and more transient weather-related patterns in the errors that they discovered. Systematic biases are more amenable to correction because they tend to recur.

Other errors arise from the chaotic nature of the atmosphere and changing climate conditions, including phenomena that may never have been observed before. Those are much harder to predict. The hope was that by looking for any patterns, at least the systematic blind spots and some of the recurring weather-related errors can be corrected. And that is in fact what happened. The researchers have demonstrated that correcting the predictable component can substantially improve forecasting skill.

And as new data becomes available, the AI can be updated and improved. It's probably not something others have not thought about, but it is an interesting direction for AI and science in general, when enough data exists or can be generated to compare simulation results with reality.

Doug Black (5:49)

Intel has disclosed additional details on its latest AI accelerator, previously known as Crescent Island, and now called the Intel GPU-D series. The new D270 offers up to 480 gigabytes of memory in partner configurations, 1.5 terabytes per second of memory bandwidth, and 1.3 petaflops of peak floating point 8 computation within a 350 watt power envelope. Intel's own card has 160 gigabytes of memory. The accelerator is designed primarily for AI inference, particularly for large models, long context applications and workloads where memory capacity and operating costs are important.