NVIDIA unveils Nemotron science stack and Nsight porting agent
NVIDIA’s SC26 software releases pair a locally deployable science-model stack with a public agent skill for profile-guided CPU-to-GPU code migration.
At SC26, NVIDIA unveiled a Nemotron for Science package and an Nsight AI Porting Agent. The releases give researchers a local deployment path for a science-model stack and a public workflow for moving supported CPU code to GPUs. NVIDIA identifies a 128GB DGX Spark as the local target for the optimized Nemotron build, while the porting skill requires a GPU and a compatible compiler toolchain.
NVIDIA’s science package combines Nemotron 3.5 Super VL, scientific evaluations, post-training recipes and LitmusBench v0.2, an open chemistry dataset with nearly 50,000 questions drawn from molecular-property and experimental drug-discovery data. NVIDIA says researchers can use these components to evaluate chemistry reasoning and adapt agents for specialized scientific tasks.
For local deployment, NVIDIA says an optimized Nemotron 3.5 Super VL build can run on a 128GB DGX Spark. The full-precision model card describes a 120-billion-parameter mixture-of-experts model with 12 billion active parameters, lists OpenMDW 1.1 as its license, and provides serving instructions for vLLM, SGLang and Docker-based tooling. The BF16 checkpoint lists a reference deployment of four H100 GPUs or two GB200 GPUs and directs DGX Spark users to an optimized NVFP4 release.
NVIDIA’s post-training recipe turns experimental input-output records into a verifiable-reward environment with NeMo Gym, then applies reinforcement learning through NeMo RL. In a drug-metabolism example with a 16,384-token output budget, NVIDIA says supervised fine-tuning raised accuracy by 12% and reinforcement-learning post-training added another 17%. The opened sources did not include an independent reproduction of those gains.
The Nsight porting skill, published in NVIDIA’s skills repository as version 6.0.0 under CC-BY-4.0 and Apache-2.0, requires a GPU, Git, the Nsight Systems command-line profiler and a target-specific compiler toolchain. It supports Fortran with OpenACC or OpenMP and C or C++ with CUDA or Kokkos. Using profiler traces, the agent plans iterative rounds of refactoring and translation before rebuilding, testing and re-profiling the application.
The porting workflow requires standalone per-function comparisons with the original implementation, along with CPU-and-GPU integration checks. Its current flavor packs do not support Python, PyTorch or JAX code. The published material also does not establish typical migration success rates or how much human intervention a production port requires.
More news

NVIDIA Introduces Four-Node DGX Spark HPC Developer Kit

NVIDIA Specifies 192-CPU Vera Rack for Scientific Computing

AMD introduces open-source HPC Assistant at SC26
