Grafy News
Try Grafy free

NVIDIA Announces Nemotron for Science, AI Agent Tools and Rubin GPU

At SC26 NVIDIA launched Nemotron for Science, an AI agent toolkit, the Rubin GPU and Vera CPUs, aiming to accelerate research with AI, quantum and simulation.

Illustration for the news story: NVIDIA Announces Nemotron for Science, AI Agent Tools and Rubin GPUSource: NVIDIA Press Room

What NVIDIA announced at SC26?

During the opening of NVIDIA’s week at SC26 in Chicago, VP of hyperscale and high‑performance computing Ian Buck unveiled a suite of new tools aimed at scientific researchers. The announcements included:

  • Nemotron for Science, a model built on the Nemotron 3.5 Super VL architecture with an open chemistry dataset and post‑training recipes for life‑science and chemistry workloads.
  • An AI agent toolkit that can plan experiments, run code, and verify results on a researcher’s own hardware.
  • The Rubin HPC GPU, part of the Vera Rubin platform, designed for science, engineering, analytics and AI.
  • An expanded Vera CPU platform for the agent loop, with early adopters such as the German Climate Computing Center (DKRZ) planning a HLRE‑5 supercomputer.
  • A joint quantum‑computing co‑design effort with IBM and Fermilab, including new error‑correction techniques and a CUDA‑Q decoding algorithm.

All of these were presented as a way to combine simulation, AI and quantum computing into a single scientific stack, according to the NVIDIA Press Room.

How AI agents are changing scientific workflows

Buck highlighted several live examples that illustrate the impact of agents on research productivity. At Argonne National Laboratory, an agent built with NVIDIA’s toolkit took a single natural‑language prompt, generated hypotheses, selected appropriate simulations, and identified two new binding sites – all without human intervention. The system runs on the researcher’s own DGX Spark, workstation, or supercomputer, keeping data and models under the user’s control.

Other showcased agents include:

  • GROMACS agent from KTH Royal Institute of Technology, which automatically repairs broken GROMACS setups. In an overnight run on a single DGX Spark it fixed 81% of test cases.
  • SETI Institute agent that, from a one‑line prompt, plotted a methanol line and then retuned for Doppler shift on its own.
  • Nsight AI Porting Agent, which ported a routine from the Weather Research and Forecasting (WRF) model to GPUs, achieving more than a 5× speed‑up using roughly 150 million tokens.

These cases demonstrate that agents are moving from novelty to practical assistants that can orchestrate complex workflows, evaluate results, and iterate without constant human supervision.

Details on Nemotron for Science and its open chemistry dataset

Nemotron for Science blends the Nemotron 3.5 Super VL open model with a curated chemistry dataset. Post‑training recipes tailor the model for life‑science and chemistry tasks, allowing researchers to run the model locally on a DGX Spark or any compatible system. Buck emphasized ownership: “You own your data. You own the model.”

The open‑chemistry dataset is meant to accelerate tasks such as molecular docking, property prediction, and reaction pathway exploration. By providing a ready‑to‑use model, NVIDIA hopes to lower the barrier for labs that lack large‑scale cloud resources while still delivering state‑of‑the‑art performance.

New hardware: Rubin HPC GPU and expanded Vera CPUs

The Rubin HPC GPU is the first GPU in the Vera Rubin platform, targeting scientific, engineering, analytics and AI workloads. Key specifications disclosed at SC26 include:

  • Support for precision modes from FP64 down to FP4.
  • Sparse‑data acceleration and seventh‑generation NVLink connectivity.
  • More than 50,000 CUDA cores – roughly three times the core count of the Hopper architecture.
  • Expected availability in 2028.

On the CPU side, agents rely heavily on the host processor to execute tool calls between model steps. NVIDIA highlighted the Vera CPU family, noting that DKRZ will install Vera CPUs in its upcoming HLRE‑5 system, joining earlier Vera deployments at the Texas Advanced Computing Center. The CPUs are positioned as the “exactly what agents need” for evaluating tool outputs, gathering results, and feeding them back into the model loop.

Quantum computing collaboration and error‑correction breakthroughs

Buck announced a new partnership involving NVIDIA, IBM and Fermilab to create a quantum‑computing co‑design platform focused on error‑correction and application research. The collaboration pairs NVIDIA DGX B200 systems with IBM quantum hardware, including the IBM Quantum Nighthawk.

A notable technical advance presented was a CUDA‑Q decoding algorithm called TCBP‑JOBS, which reduces decoding latency from about 300 µs to 10 µs. The same algorithm cuts the time required for a quantum error‑correction study from 12 weeks down to 45 minutes. Additionally, a simulation on Jülich’s JUPITER supercomputer demonstrated that the magic‑state cultivation technique can lower the number of operations needed for a logical step by a factor of ten, a result that will be integrated into CUDA‑Q.

Practical implications for researchers and developers

The combined hardware and software announcements suggest a shift toward tightly integrated, end‑to‑end scientific pipelines:

  • Local execution with data sovereignty – Researchers can keep proprietary data on‑premise while leveraging powerful AI models.
  • Agent‑driven automation – Routine tasks such as simulation setup, parameter sweeps, and result validation can be delegated to agents, freeing scientists to focus on interpretation.
  • Scalable performance – The Rubin GPU’s massive core count and the Vera CPUs’ suitability for the agent loop promise substantial speed‑ups for both traditional HPC codes and AI‑augmented workloads.
  • Quantum‑ready workflows – By exposing error‑correction primitives through CUDA‑Q, developers can begin integrating quantum steps into larger HPC pipelines today.

For developers, the new Nsight AI Porting Agent provides a concrete example of how to migrate legacy code to GPUs with minimal manual effort. The 5× speed‑up reported for the WRF model demonstrates the potential ROI of adopting NVIDIA’s agent‑based tooling.

Grafy perspective: experimenting with AI agents

While Grafy does not yet host NVIDIA’s proprietary models, creators and researchers can explore similar agent workflows using existing tools in the Grafy ecosystem. The Grafy Chat app lets users interact with multiple open‑source models side‑by‑side, making it a convenient sandbox for drafting prompts and testing how an agent might plan an experiment. For developers interested in building or extending agents, Grafy Code offers an AI‑assisted coding environment where you can prototype tool‑calling logic, run code snippets, and iterate quickly.

If you have a PDF of a scientific paper you want to feed into an agent, the Grafy Reader can ingest the document, let you ask questions, and retrieve relevant passages – a useful step when constructing a knowledge base for an agent. By chaining these apps—using Reader to load literature, Chat to formulate hypotheses, and Code to automate simulation runs—users can approximate the end‑to‑end workflow that NVIDIA showcased, all within the Grafy workspace.


Frequently asked questions

How does Nemotron for Science differ from the standard Nemotron model?

Nemotron for Science builds on the open‑source Nemotron 3.5 Super VL architecture but adds a curated chemistry dataset and post‑training recipes specifically tuned for life‑science and chemistry tasks. This specialization lets researchers run the model locally on their own hardware while keeping data private.

What performance advantage does the Rubin HPC GPU claim over previous generations?

According to NVIDIA, the Rubin GPU contains more than 50,000 CUDA cores—about three times the core count of the Hopper architecture—and supports a wide precision range from FP64 to FP4, plus sparse‑data acceleration. These features are intended to accelerate both traditional HPC simulations and AI workloads.

Can I prototype the AI agent workflow described by NVIDIA using Grafy tools?

Yes. Grafy Chat can be used to experiment with prompt‑driven agents, while Grafy Code lets you write and run the tooling logic that an agent would invoke. For literature‑driven agents, Grafy Reader can ingest PDFs and provide searchable context. Together, these apps let you build a proof‑of‑concept of the agent pipeline without needing NVIDIA‑specific hardware.


Source: NVIDIA Press Room — Ian Buck Kicks Off NVIDIA’s Week at SC26 as Science Enters the Age of Agents