AlphaGazette
Menu

AI

Buck unveils Nemotron for Science so research agents can stay on lab systems

The open package pairs Nemotron 3.5 Super VL with a chemistry dataset and life-science recipes that labs can run themselves. In the same SC26 address, Nvidia detailed a quantum codesign effort with IBM and Fermilab.

2 min read 3 sources AI-generated

Ian Buck, vice president for hyperscale and high-performance computing at NVIDIA, began the company's SC26 week in Chicago on Monday with software that lets an agent plan an experiment, execute code and verify results on machines a research group already runs. In that address he introduced Nemotron for Science, a wider Vera CPU platform built for high-performance computing, and a quantum collaboration with IBM and Fermi National Accelerator Laboratory. The talk also covered the Rubin HPC GPU.

Agents that keep data and weights local

Nemotron for Science joins the open Nemotron 3.5 Super VL model to a newly released open chemistry dataset and to post-training recipes for chemistry and the life sciences. A lab can run the package on a DGX Spark, a workstation or a supercomputer it operates. Buck told the audience, "You own your data. You own the model."

Examples he gave included an Argonne National Laboratory agent that, from one prompt, planned hypotheses, tested them, decided which simulations should run and found two new binding sites. A GROMACS agent written at KTH Royal Institute of Technology repairs failed setups on its own; on a single DGX Spark, one overnight pass put right 81% of its test cases. At the Allen Telescope Array, a SETI Institute agent accepted a one-line prompt, drew the methanol spectral line and then corrected for Doppler shift without a further instruction. Buck said NVIDIA's new Nsight AI Porting Agent carried a Weather Research and Forecasting routine onto GPUs, lifting speed by more than 5x and consuming about 150 million tokens.

Library speedups and the quantum codesign

NVIDIA is extending CUDA-X with cuPhoton, cuPotential and cuSpa so agents can search large scientific datasets. In the cuPhoton Toolkit, cuPhoton.xBeam correlates radio data from the Event Horizon Telescope: a job that previously occupied over a month now completes in eight minutes. The company said the module correlated jets known as J1924 and 3C279, and that loading, processing and visualization can run across distributed multi-GPU machines or on a compact DGX Spark. Pacific Fusion used cuPhoton.xLens to extract findings while a plasma simulation is still underway, and a Nemotron-powered agent on that same system can help steer the run. Buck put NVIDIA's CUDA-X catalog at more than 1,000 libraries. Agents, he said, rely on CPUs to make tool calls between model steps, and that is the workload he tied to the broader Vera CPU platform.

Nvidia said the work with IBM and the US Department of Energy's Fermi National Accelerator Laboratory will examine modular quantum machines, error correction and uses in science. The Next Web reported that a DGX B200 is at the platform's core. Buck's framing was that simulation stays the base of scientific computing, with AI and quantum beside it.