NVIDIA announces Rubin HPC GPU for 2028


NVIDIA says the planned chip will balance data movement with computation for scientific simulations and data processing, giving researchers and engineers a roadmap for future hardware.
NVIDIA announced Rubin HPC at SC26 on 16 November, a graphics processing unit (GPU) designed for simulation and data processing in science, engineering and analytics. The chip is part of the company’s Vera Rubin platform and is expected to be available in 2028.
Scientific simulations help researchers test materials, molecules and designs before physical experiments or prototypes. NVIDIA says faster simulations can give researchers feedback sooner as they decide what to test next.
In its announcement, NVIDIA said scientific applications often gather scattered values from large datasets and repeatedly calculate results. Rubin HPC is designed to increase data movement alongside computing capacity, so its cores spend less time waiting for data.
The company says the chip will have more than 50,000 CUDA computing cores, a feature for supplying them with scattered data, and seventh-generation NVLink connections between GPUs.
NVIDIA also announced a rack-scale platform connecting 192 Vera central processing units (CPUs) through Quantum-X800 InfiniBand networking. It is intended for workloads including climate modeling and molecular dynamics. NVIDIA says Vera combines custom cores, high-bandwidth memory and networking to move large datasets and sustain simulation performance as workloads grow.
In NVIDIA’s tests on OpenFOAM’s motorBike Large airflow simulation, Vera delivered 1.9x the performance and 2.4x the performance per watt of x86 systems. For the Weather Research and Forecasting model, the company projects 1.6x the performance and 2x the energy efficiency compared with x86.
NVIDIA says the German Climate Computing Center selected Vera CPUs, GB200 Grace Blackwell systems and Quantum-2 InfiniBand networking for its HLRE-5 supercomputer. Bull is building the system, which NVIDIA expects to be available in early 2028 for climate simulation and superintelligence (SI) research.
NVIDIA introduced Nemotron for Science, combining the Nemotron 3.5 Super VL model, an open chemistry dataset and post-training recipes for adapting the model to specialist tasks. The company says the LitmusBench v0.2 dataset contains nearly 50,000 chemistry questions drawn from molecular properties and experimental drug-discovery data for training and evaluating language models on chemistry reasoning. The model and dataset are available on Hugging Face.
The company also expanded its CUDA-X software libraries, including cuPhoton, cuPotential and cuSpa, for research ranging from astrophysics and materials science to single-cell biology.
NVIDIA announced plans with IBM and the US Department of Energy’s Fermi National Accelerator Laboratory for a platform to evaluate quantum error correction, applications combining quantum and conventional computing, and designs linking multiple quantum processors to conventional systems.
Separately, NVIDIA released the TCBP-JOBS algorithm through CUDA-Q QEC 0.9.0. It is a decoder: software that interprets measurements to identify likely errors for correction. NVIDIA says it can decode quantum error-correction research workloads up to 13,000x faster than existing leading algorithms.
NVIDIA also expanded quantum-processor support in Slurm, software used to schedule supercomputing work. Previously, integrating each processor or center posed a new software challenge. NVIDIA says hardware builders can reuse integrations across deployments, while centers can add processors to existing scheduling workflows. Slurm 26.11 is expected later in November 2026.
NVIDIA says cloud provider Lambda is using its Photonics optical switches to connect GB300 NVL72 racks containing more than 10,000 GPUs.
In the MLPerf Training benchmark, which measures how quickly systems train SI models to a target accuracy, NVIDIA says its Vera Rubin NVL72 system, distinct from the announced Rubin HPC GPU, delivered up to 2x the performance per GPU of GB300 NVL72.
NVIDIA also says a provisional MLPerf Endpoints result for the Kimi K3 model showed Vera Rubin NVL72 delivering more than 10x the output per watt of GB300 NVL72, at a shared minimum average response rate of 50 tokens per second per user. Tokens are units of model output. The Vera Rubin result awaits MLCommons verification and could change after peer review.
NVIDIA expects Vera CPU racks to be available from Dell Technologies, HPE, Lenovo and Supermicro beginning in early 2027.
La Poste confirmed over the weekend that a September attack on Maileva exposed recipients' names and postal addresses. Customers have described additional information that may have been exposed.
3 min

Liquid AI says its decision models add audio input and five task types. It says d1.2-ultra scored 75 to lead the Artificial Analysis Decision Index v2.1.
3 min

The interim final rule requires licenses to reexport advanced superintelligence (SI) chips to Indonesia, Malaysia, Singapore, Thailand or Brazil.
1 min