NVIDIA announces Rubin HPC GPU and Vera CPU for 2028 scientific computing

NVIDIA’s Rubin HPC GPU features 50,000+ CUDA cores—three times the count of the Hopper architecture—and is scheduled for 2028 availability. On 17 November 2026, Ian Buck announced these specs along with the Vera CPU platform and Nemotron for Science at SC26. If you already own a Hopper-based GPU or are currently using CUDA-X libraries for simulation, these new Rubin and Vera platforms represent your next hardware cycle rather than an immediate upgrade.

Is this for you?

If you are a supercomputer architect, the Rubin HPC GPU and Vera CPU represent the hardware cycle for your 2028 deployments.

  • If you are a supercomputer system architect, the Rubin HPC GPU and Vera CPU are the primary components for your 2028 hardware procurement contracts.
  • If you are a quantum computing researcher, the TCBP-JOBS algorithm is the key for reducing error-correction research computation from 12 weeks to 45 minutes.
  • If you are an HPC software developer, the Vera CPU platform is the necessary processing power for the logic-heavy agent loop.
  • Skip this if you are a gaming enthusiast looking for consumer GPU launches as this focuses on enterprise-scale science hardware.

What you can do now

  1. If you are a supercomputer system architect, verify that the HLRE-5 supercomputer hardware delivery schedule aligns with the 2028 Rubin HPC GPU availability.
  2. If you are an HPC software developer, check the CUDA-Q library for the inclusion of the TCBP-JOBS decoding algorithm.
  3. If you are an HPC software developer, identify which of the 1,000+ CUDA-X libraries are required to run agents on the Rubin HPC GPU.

Key figures

Rubin CUDA cores
50,000+ (3x Hopper), announced 17 November 2026
Rubin availability
Expected 2028
TCBP-JOBS reduction
12 weeks to 45 minutes
Vera CPU deployment
HLRE-5 supercomputer, early 2028

Key facts

Rubin CUDA cores
50,000+ (3x Hopper), announced 17 November 2026
Rubin availability
Expected 2028
TCBP-JOBS reduction
12 weeks to 45 minutes
Vera CPU deployment
HLRE-5 supercomputer, early 2028

Rubin HPC GPU: The next generation of scientific computing

The Rubin HPC GPU is a new processor in the Vera Rubin platform designed for science, engineering, and AI. It supports a wide range of precisions from FP64 to FP4 and incorporates sparse-data acceleration—hardware optimization for processing non-contiguous or irregular data structures. (FP64 to FP4: different levels of numerical precision used in computer calculations for scientific and artificial intelligence tasks.) The architecture connects via seventh-generation NVIDIA NVLink for high-bandwidth interconnects. (NVLink: a high-bandwidth interconnect technology used to connect multiple graphics processing units together.)

NVIDIA says the Rubin GPU provides about three times as many CUDA cores

NVIDIA says the Rubin GPU provides about three times as many CUDA cores as the preceding Hopper architecture. This significant core count increase, combined with wide precision support, makes Rubin suitable for both high-precision scientific simulation and low-precision AI inference. The hardware is scheduled for release in 2028.

Vera CPU platform and the agent loop

The Vera CPU platform provides the necessary processing power for the “agent loop,” a workflow where AI agents perform tool calls between model steps. (agent loop: a workflow where artificial intelligence agents perform tool calls between various model steps.) Buck explained that agents require a CPU to run, evaluate, and gather results from various tools before the next model step occurs. The Vera CPU is positioned to handle these logic-heavy operations, allowing the GPU to focus on primary simulation tasks.

The German Climate Computing Center (DKRZ) has already selected Vera CPUs for its HLRE-5 supercomputer, which is expected in early 2028. This selection confirms the platform’s role as a foundational component for the next generation of supercomputing infrastructure.

Nemotron for Science and agentic workflows

Nemotron for Science combines the Nemotron 3.5 Super VL open model with a new open chemistry dataset and post-training recipes—specific sets of instructions for fine-tuning models—for life sciences. (Nemotron 3.5 Super VL: an open model used in combination with chemistry datasets for life sciences research.) The model is designed to be run locally on a DGX Spark, a workstation, or a supercomputer, helping ensure users own both their data and the model.

NVIDIA demonstrated several practical applications for these agents.

NVIDIA demonstrated several practical applications for these agents. An agent built with the toolkit at Argonne National Laboratory took a single prompt to plan and test hypotheses, identifying two new binding sites. Additionally, an agent from the KTH Royal Institute of Technology fixed 81% of test cases for broken GROMACS setups in a single overnight run on a DGX Spark. (GROMACS: a software package used for molecular dynamics simulations and research.)

Quantum computing breakthroughs

NVIDIA, IBM, and Fermilab announced a quantum computing co-design platform to support error-correction and applications research. A simulation on Jülich’s JUPITER supercomputer showed that magic state cultivation—a quantum error correction technique to reduce operations for logical steps—can cut necessary operations tenfold. This work is being integrated into the CUDA-Q library.

Buck also introduced the TCBP-JOBS algorithm

Buck also introduced the TCBP-JOBS algorithm, which significantly reduces quantum decoding latency. The goal is to cut decoding latency from 300 microseconds to 10 microseconds. The TCBP-JOBS algorithm enables rapid iteration in quantum research by slashing the time required for error-correction research computation from 12 weeks to 45 minutes.

Photonics and interconnect reliability

NVIDIA showed one of the first deployments of NVIDIA Photonics co-packaged optics—integrated light-based networking—switches at Lambda. These switches use co-packaged optics to improve reliability in high-scale computing environments. NVIDIA said the switches are 15x more reliable on bit errors and link flaps compared to previous solutions.

Who this is for: Supercomputer system architects planning for 2028

Supercomputer system architects planning for 2028 are for those who need to specify hardware for next-generation deployments like the HLRE-5. These architects should verify the hardware delivery schedule for the HLRE-5 supercomputer to align with the Rubin HPC GPU availability in 2028.

Quantum computing researchers seeking decoding improvements are for those looking to reduce error-correction overhead. These researchers should monitor the CUDA-Q library for the inclusion of the TCBP-JOBS decoding algorithm.

HPC software developers implementing agents are for those who need to integrate AI agents into scientific workflows. They should identify which of the 1,000+ CUDA-X libraries are required to run agents on the Rubin HPC GPU.

Data center network engineers managing high-scale networks are for those who need reliable interconnects. They should check the compatibility of the seventh-generation NVIDIA NVLink with existing networking infrastructure.

Who this is not for: Gaming enthusiasts looking for consumer GPU launches

Gaming enthusiasts looking for consumer GPU launches can ignore this section as the focus is on high-performance computing (HPC) and enterprise-scale science hardware. These readers should instead follow consumer-grade GPU launch news for gaming-specific hardware.

A research scenario for Dr. Amara

Dr. Amara is a researcher at a national laboratory needing to identify new binding sites in a protein. Dr. Amara uses an agent built with the NVIDIA toolkit, providing a single prompt that allows the agent to plan and test hypotheses, identify simulations to run, and identify two new binding sites.

The research is completed significantly faster using an agent-driven flow rather than manual simulation planning. If Dr. Amara did not have the agent toolkit, he would have to manually select simulations and check results, a process that is less efficient than the automated agent loop.

Practical applications for the research stack

Application Target User
Reduce quantum error-correction computation time from weeks to minutes Quantum computing researchers
Automatically repair broken GROMACS setups Molecular dynamics researchers
Port WRF model routines to GPUs with a 5x speedup Meteorological researchers
Run life science models locally without sharing data Privacy-conscious research teams

Nsight AI Porting Agent replaces manual WRF model porting

The Nsight AI Porting Agent replaces manual WRF model porting, which it claims achieves a 5x speedup using approximately 150 million tokens. (WRF model: a numerical model used by researchers for predicting weather and climate.) The TCBP-JOBS algorithm replaces the 12-week quantum decoding computation with a 45-minute process, significantly reducing the research cycle.

What is not yet known: The source did not state when the TCBP-JOBS algorithm

The source did not state when the TCBP-JOBS algorithm will be available in the CUDA-Q library. Additionally, the source did not confirm if the Rubin HPC GPU is compatible with IBM Quantum Nighthawk systems, nor did it disclose the specific configuration of the DGX Spark used for running Nemotron for Science. Finally, the specific limitations of the TCBP-JOBS algorithm were not disclosed.

Rubin HPC GPU: what to check, compare and watch

Hopper gpu
The Rubin HPC GPU is a successor to the Hopper architecture with significantly more CUDA cores. Cuda core count: three times as many as Hopper. Supercomputer system architects: compare the total CUDA core count of the Rubin HPC GPU against the current Hopper GPU units.
Hlre-5 deployment
The HLRE-5 supercomputer is scheduled for deployment in early 2028, aligning with the Rubin HPC GPU release. Deployment date: early 2028. Supercomputer system architects: verify the hardware delivery schedule for the HLRE-5 supercomputer to align with the Rubin HPC GPU availability.
Procurement contract
The Rubin HPC GPU will be a primary component in hardware procurement contracts for large-scale systems. Supercomputer system architects: update the technical specifications section in the hardware procurement contract to include Rubin HPC GPU requirements.
Gpu cluster
Researchers may use a GPU cluster as an alternative to the Rubin HPC GPU for quantum simulation. Quantum computing researchers: compare the Rubin HPC GPU performance against a standard GPU cluster for magic state cultivation simulations.
Veras cpu
Agents require Vera CPUs to run tool calls between steps when using the Rubin HPC GPU. HPC software developers: verify that the Vera CPU platform meets the processing requirements for the agent tool call loop.
Cuda-q decoding algorithm
The TCBP-JOBS algorithm reduces error-correction computation time for workflows utilizing the Rubin HPC GPU. Computation time: 45 minutes. HPC software developers: check the reduction in computation time from 12 weeks to 45 minutes using the TCBP-JOBS algorithm.

Rubin HPC GPU: questions the original report leaves open

What is the main finding, in numbers?
The original report did not say what the main finding was.
Who carried out the study, and who paid for it?
The original report did not say who carried out the study or who paid for it.
How large was the sample, and what method and time period were used?
The original report did not say the sample size, method, or time period.
What are the limitations?
The original report did not say what the limitations are.
What should practitioners change because of it?
The original report did not say what practitioners should change.
When will the TCBP-JOBS algorithm be available in CUDA-Q?
The original report did not say when the TCBP-JOBS algorithm will be available.
Is the Rubin HPC GPU compatible with IBM Quantum Nighthawk systems?
The original report did not say if the Rubin HPC GPU is compatible with IBM Quantum Nighthawk systems.
What is the configuration of the DGX Spark for running Nemotron for Science?
The original report did not say what the configuration of the DGX Spark is.

Key terms

Sparse-data acceleration
Faster irregular data processing—hardware optimization for non-contiguous data structures.
Co-packaged optics
Integrated light-based networking—optical interconnects integrated into the same package as the processor or switch.
Magic state cultivation
Quantum error correction technique—a method to reduce operations for logical steps in quantum computing.
Post-training recipes
Custom training instructions—specific sets of instructions for fine-tuning models after initial training.

The move toward agent-mediated workflows signals a shift where the balance between CPU

The move toward agent-mediated workflows signals a shift where the balance between CPU and GPU becomes critical; the CPU must handle the logic of the “tool call” while the GPU handles the heavy lifting of simulation. This transition, alongside the 2028 roadmap for Rubin and Vera, establishes the next generation of scientific infrastructure. While NVIDIA leads this rollout, competitors like AMD continue to develop their own HPC offerings, meaning the final selection will depend on specific software library support and research-specific workflows.

Reported by NVIDIA on 17 November 2026: Ian Buck Kicks Off NVIDIA’s Week at SC26 as Science Enters the Age of Agents.

Related articles

ExtremeSpec is reader-supported: some links on this page are affiliate links, and as an Amazon Associate ExtremeSpec earns from qualifying purchases at no extra cost to you. Learn more.

Similar Posts