Models
Nvidia announces the Rubin HPC GPU with more than 50,000 CUDA cores for 2028, and the GR00T 2 robot model trained on about 402,730 hours of human video
Nvidia unveiled Rubin HPC (50,000+ CUDA cores, due 2028) at SC26 and GR00T 2 (not “N2”), a robot model trained on about 402,730 hours of human video.
Nvidia presented two products in the past week. At the SC26 supercomputing conference on November 16 it announced Rubin HPC, a GPU with more than 50,000 CUDA cores that is due in 2028. Earlier, on November 9, it announced a robot model called GR00T 2, not “N2” as sometimes written, trained on roughly 402,730 hours of first-person human video. Several details, such as memory, power and price for the chip, are not disclosed.
The essentials
- Rubin HPC: announced at SC26 with more than 50,000 CUDA cores, a record for Nvidia, plus sparse-data acceleration and seventh-generation NVLink. It arrives in 2028.
- For scale: today’s Rubin GPU has 224 SMs and 28,672 CUDA cores, so the HPC part would be nearly twice that. The layout, likely multi-die, is not disclosed.
- GR00T 2: a “world action model” built on Cosmos, announced November 9 at CoRL in Austin. The name is GR00T 2.
- The data: about 402,730 hours of first-person human video, per one source, plus 10,184 hours of handheld-gripper demos and 23,500 hours of robot teleoperation.
What Rubin HPC is and is not
It is a data-center GPU aimed at scientific computing, with the core count as the headline. Memory capacity, double-precision (FP64) performance, power draw and price are not disclosed in the coverage we read. The 2028 date is a plan, not a shipping product.
What GR00T 2 claims
Nvidia’s own results, as relayed by The Neuron, include 40 of 50 tasks completed at least once without task-specific demonstrations, and a 95% success rate with five demonstrations after a 30-minute fine-tune on one RTX 5090. These are self-reported. No parameter count or license is published, and code is listed for December 2026, so it is not yet a released open model.
What is not established
- Rubin HPC memory, FP64 performance, power and price: not disclosed.
- The chip’s die layout: not disclosed.
- GR00T 2 parameter count and license: not published.
- Independent replication of the GR00T 2 results: none yet.
- The HPCwire comparison with Blackwell: page not read.
Summary table
| Point | What is stated as of November 17, 2026 | Source |
|---|---|---|
| Rubin HPC CUDA cores | More than 50,000 (single GPU) | Tom’s Hardware |
| Rubin HPC timing | 2028 | Guru3D |
| Today’s Rubin | 224 SMs, 28,672 CUDA cores | Shattered |
| GR00T 2 human video | About 402,730 hours | The Neuron |
| GR00T 2 code | Listed for December 2026 | Nvidia Research |
FAQ
Is the robot model called GR00T N2?
No. The sources we read call it GR00T 2.
When can people buy Rubin HPC?
Nvidia says it arrives in 2028. No price or detailed specifications have been given.
Are the GR00T 2 results independent?
No. They are Nvidia’s own reported figures, and the code is not yet released.
Method and limits
Checked on November 17, 2026. We read the pages or search excerpts of Tom’s Hardware, Guru3D, Shattered, The Neuron and Nvidia Research. The HPCwire page could not be opened. GR00T 2 benchmark numbers are self-reported.
Sources
- Tom’s Hardware: Nvidia announces Rubin HPC GPU with record number of CUDA cores (verified 2026-11-17)
- Guru3D: Nvidia Rubin HPC GPU brings more than 50,000 CUDA cores to scientific computing (verified 2026-11-17)
- Shattered: Nvidia Rubin HPC, 50,000 CUDA cores, due 2028, SC26 (verified 2026-11-17)
- The Neuron: Nvidia GR00T 2 learns robot tasks from human video (verified 2026-11-17)
- Nvidia Research: GR00T 2 (verified 2026-11-17)
Spotted an error in this article? Report it · Our corrections policy