NVIDIA unveils AI agents and new hardware for scientific research
NVIDIA introduced tools that allow AI agents to plan experiments, run code, and manage complex scientific simulations autonomously.
The short version
- Nemotron for Science combines a 3.5 Super VL open model with new chemistry and life science datasets.
- AI agents can now autonomously repair broken setups, plot telescope data, and port weather models to GPUs.
- The Rubin HPC GPU is expected in 2028 and features over 50,000 CUDA cores.
- NVIDIA is partnering with IBM and Fermilab on a quantum computing co-design platform for error correction.
- A new CUDA-Q decoding algorithm can reduce error-correction research time from 12 weeks to 45 minutes.
Why it matters
This shifts the role of researchers from manually running every step of an experiment to overseeing autonomous agents that handle the heavy lifting. It accelerates the pace of discovery in fields like chemistry, climate science, and quantum computing.
The rise of scientific agents
NVIDIA's Ian Buck opened the SC26 conference by introducing a new era of AI agents designed specifically for scientific research. Unlike standard chatbots, these agents can plan experiments, execute code, and verify results on a researcher's own local systems. This approach allows scientists to interact with supercomputers and instruments more naturally through high-level instructions.
Buck explained that while simulation remains the core foundation of scientific computing, AI and quantum technologies are now working alongside it. The goal is to create a new scientific stack where agents handle the repetitive and complex logic of data processing. This allows researchers to focus on high-level hypothesis generation rather than manual data manipulation.
The company showcased several real-world examples of these agents in action across different fields. At the Argonne National Laboratory, an agent took a single prompt to identify new binding sites by planning and testing various hypotheses. In other cases, agents have been used to repair broken software setups overnight or plot telescope data with minimal human input.
Nemotron for Science and local AI
A major software release highlighted during the event is Nemotron for Science. This tool combines the Nemotron 3.5 Super VL open model with new datasets specifically curated for chemistry and the life sciences. It includes post-training recipes that help the model understand complex biological and chemical structures.
One of the key advantages of this release is that researchers can run the model locally on their own hardware. This includes workstations, supercomputers, or specialized DGX Spark systems. By keeping the model on-site, researchers maintain full ownership of their sensitive data while still benefiting from advanced AI capabilities.
The practical utility of these models is already being tested in diverse environments. For instance, an agent built with NVIDIA's toolkit helped researchers at the SETI Institute retune telescope data for Doppler shifts automatically. These tools are designed to bridge the gap between raw data collection and actionable scientific insights.
New hardware for the agent loop
To support these agents, NVIDIA is introducing new hardware designed for the specific demands of the 'agent loop.' Because agents need to run tool calls, evaluate results, and perform operations between model steps, they rely heavily on CPU performance. The expanded Vera CPU platform is being positioned as a key component for these types of workloads.
The German Climate Computing Center has already selected Vera CPUs for its upcoming HLRE-5 supercomputer, which is expected to launch in early 2028. These processors are intended to handle the logic and operations that occur between the different steps of an AI agent's workflow. This ensures that the transition from thinking to acting is as fast as possible.
In addition to CPUs, NVIDIA announced the Rubin HPC GPU, which is built for science, engineering, and analytics. This new GPU is expected in 2028 and features more than 50,000 CUDA cores, which is about three times as many as the previous Hopper architecture. It supports a wide range of precisions and includes new acceleration for sparse data.
Quantum co-design and networking
NVIDIA is also moving into the quantum space through a new collaboration with IBM and the Fermi National Accelerator Laboratory. They are developing a quantum computing co-design platform to support error correction and research applications. This involves pairing NVIDIA DGX B200 systems with IBM quantum computers, including those powered by IBM Quantum Nighthawk.
A significant part of this effort focuses on reducing latency and computation time for quantum decoding. A new CUDA-Q decoding algorithm, called TCBP-JOBS, was shown to cut error-correction research computation from 12 weeks down to just 45 minutes. Additionally, simulations on the JUPITER supercomputer showed that a technique called magic state cultivation could cut the operations needed for one logical step tenfold.
Finally, the company demonstrated new networking hardware to support these massive scientific workloads. At Lambda, NVIDIA showed one of the first deployments of Photonics co-packaged optics switches. These switches are reported to be 15x more reliable regarding bit errors and link flaps, which is critical for maintaining stable connections in large-scale supercomputing clusters.
What to watch
- The rollout of the Vera CPU platform in 2028.
- How researchers at Argonne National Laboratory scale their binding site discovery using agents.
- The integration of magic state cultivation into the CUDA-Q library.
Try it
Look up the Nemotron for Science model on NVIDIA's website to see the available chemistry and life science datasets.
Sources and links
First reported by blogs.nvidia.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.