Nvidia's Rubin HPC GPU targets science, and it won't ship until 2028

7 Sources

Share

NVIDIA announced its largest disclosed CUDA core count yet for a single GPU, aimed at scientific simulation. The specs buyers need are missing, and the chip is expected in 2028.

NVIDIA has announced the Rubin HPC GPU, a chip aimed at scientific simulation and engineering rather than AI alone, and it carries the largest CUDA core count NVIDIA has disclosed for a single GPU: more than 50,000

5

2

1

.

Source: SiliconANGLE

Source: SiliconANGLE

The chip debuted at the SC26 supercomputing conference in Chicago on November 16

5

2

1

, and it is not expected to be available until 2028

2

4

.

Source: Analytics Insight

Source: Analytics Insight

Built for simulation

NVIDIA's argument is that faster simulation lets AI agents test ideas, check the results and steer the next run. Scientists and engineers can then screen more candidate materials, molecules and designs. "For agents tackling science and engineering, faster simulation means faster discovery," said Ian Buck, the company's vice president of hyperscale and high-performance computing

5

.

The design reflects that pitch. Rubin HPC uses seventh-generation NVLink to connect GPUs to each other and includes a sparse data acceleration feature whose job is keeping all those cores supplied with data

5

3

. For scale, the current flagship Rubin GPU has 28,672 CUDA cores. That would be at least 74% more cores, a jump worked out by TechSpot rather than claimed by NVIDIA

1

.

Source: Guru3D

Source: Guru3D

What NVIDIA left out

NVIDIA said nothing about memory type or capacity, memory bandwidth, board power, double-precision performance, configurations, price or how the chip compares to anything else

3

1

4

. Double-precision arithmetic, the math traditional scientific computing runs on, is exactly where the answer is missing. That absence matters because AMD used the same conference to promote an Instinct MI400 variant with record double-precision numbers. It is a field NVIDIA has largely sat out while it concentrated on AI

5

1

.

Even the core count has an asterisk. One report's reading of NVIDIA's released image points to two GPUs joined by NVLink rather than one, which would change what 50,000 cores means. NVIDIA has not confirmed the layout, and its own description stops at the single-GPU count

5

1

4

. None of those numbers can be checked yet. Until independent testing exists, the gains depend on model size, memory capacity, libraries and how well each workload fits the GPU

3

.

The software ships first

What researchers can use today is software. NVIDIA debuted more than half a dozen tools alongside the chip

2

6

. cuPhoton.xBeam, per NVIDIA, cut an Event Horizon Telescope analysis from a month-long task to eight minutes in one test

2

. cuSpa, an open-source tool for calculating cell shapes, ran 900-fold over CPUs in one test at Memorial Sloan Kettering Cancer Center, again by NVIDIA's account

2

6

. There is also a Nemotron for Science bundle built on a 124-billion-parameter model trained on roughly 50,000 chemistry questions, plus new CUDA-Q quantum building blocks

2

6

. Buck framed the software push the same way, saying Nemotron for Science supplies open models, data and post-training recipes for agents that put CUDA-X tools to work

2

.

Alongside all of it, NVIDIA is teaming with IBM and the US Energy Department's Fermilab on a modular quantum codesign platform. The goal is to pair DGX B200 systems with IBM quantum hardware in one codesign environment

2

.

The question one report raises, citing the cancelled Rubin CPX, is whether Rubin HPC reaches market at all, and nothing NVIDIA said at SC26 settles it

5

.

© 2026 TheOutpost.AI All rights reserved