systems

Tireless AI agents could help scientists do research humans alone can't, says GPU giant

Nvidia is pushing agentic AI for scientific computing, and says that this requires a new scientific computing stack, which the GPU giant is ready to deliver, of course.

At the ISC High Performance 2026 event in Hamburg, Germany, Nvidia is lauding its own achievements in supercomputing, highlighting just how many of the world’s top compute clusters use its hardware these days.

But just as agentic AI has become this year’s buzzword in the machine intelligence industry, so the GPU slinger is pushing it as the next big thing for supercomputers and their research programs, driven by its next-gen Vera Rubin platform and new software tools.

“We are currently witnessing a massive inflection point with agentic AI. AI is shifting from a tool that simply answers questions to an autonomous system that executes complex tasks,” Nvidia’s senior director of HPC and AI Factory Solutions Dion Harris told the media in a briefing.

The Mission and Vision systems at the Los Alamos National Laboratory (LANL) in the US will be the world's first agentic AI supercomputers when they come online, he says.

A new scientific computing stack connects agents, simulation, and AI together to accelerate the next generation of scientific discovery, Harris claims.

“Scientists leverage agentic AI co-scientists that call simulators and surrogate models alongside tools and applications, to do everything from planned experiments to write code to run the simulations to simulations and AI and data analytics converging into one single workflow,” he explained.

This requires an incredible amount of compute, memory, and networking, which, in Nvidia’s eyes, means supercomputers built on its Vera Rubin and Grace Blackwell platforms, plus Quantum InfiniBand networking, and new software for accelerating discovery.

The latter comprises ALCHEMI, DAQIRI, and cuPhoton. The first is described by Nvidia as a domain-specific toolkit for chemical and material discoveries, using the BGR microservice for simulating millions of molecules and structures.

DAQIRI is designed for the next-generation scientific instruments, connecting sensors directly to real-time AI inference points, Harris says. “At CERN's ATLAS experiment, less than 2 percent of collision data can typically be stored. DAQIRI introduces a GPU accelerated AI trigger pipeline allowing FPGAs to handle low latency routing while GPUs run deep learning models to ensure we learn from significantly more data,” he explained.

Finally, cuPhoton is built to process petabytes of camera and telescope data to help scientists analyze massive cosmic data sets in minutes rather than months.

“In testing with 32 Grace Blackwell superchips simulating data from the Rubin Observatory, cuPhoton loaded and read images 15,000 times faster and accelerated signal processing and analysis by up to 8,000 times,” Harris claimed.

But Nvidia is pitching its next-gen silicon as the platform for agentic supercomputing. Due to be available in Q4 this year, the Vera Rubin NVL rack will cram in up to 144 GPUs per rack, and deliver 5 petaFLOPS of FP64 floating-point performance.

Because many high-performance computing workloads are often bound by memory performance, Vera Rubin increases memory bandwidth by 2.8 times compared to Blackwell, Harris says, using 41 TB of HBM4 memory per rack to achieve three petabytes per second of bandwidth.

Systems that are getting Vera Rubin include the Mission and Vision systems at LANL. These stack up to 2,160 Rubin GPUs plus 1,080 Vera CPUs, in the case of Mission, while Vision has a more modest 1,298 Rubins and 648 Veras.

“Then there's Veritas, which is being announced at ISC, which deploys 576 Rubin GPUs, along with 288 Vera CPUs,” Harris says.

We asked Nvidia what the purpose is of embedding agentic AI into scientific computing, much of which is about research driven by human curiosity.

“Agentic AI, or in fact any AI, is not required to do science,” Harris told The Register.

“But Nvidia believe agentic AI is already emerging as a powerful tool to do science at a scale that isn’t possible when human scientists alone drive the process. Agents don’t need to sleep, or eat, or take breaks.  They can consume thousands or millions of technical papers and remember the details, and in some cases, they benefit from PhD-level understanding across diverse fields from astrophysics to zoology,” he said.

Nvidia’s vision is that human scientists will have a team of agents running around the clock, able to do investigations they couldn’t themselves perform.

“But agents require foundation models, LLMs, and connections to data and tools to perform science.  They run on CPUs, but access tools, many of which need GPUs to run at maximum performance and efficiency,” Harris added.

Nvidia claims that Europe is now a hotspot for HPC, with 35 new supercomputers brought online in the past year, all using Nvidia tech. These include Jupiter, Europe’s exascale system, MareNostrum 5 at the Barcelona supercomputing center, Bavaria AI's Blue Swan, HammerHAI at the University of Stuttgart, and Italy’s CINECA. ®