Nvidia's Rubin HPC chip packs 50,000+ CUDA cores, ships in 2028
Nvidia unveiled Rubin HPC at SC26 on Nov. 16, a scientific-computing GPU headed for 2028 with more than 50,000 CUDA cores and seventh-generation NVLink. The company pitches it at climate, genomics and financial risk work, but stayed quiet on memory, double-precision performance and benchmarks.
- 2 sources report it: thedefiant.io, blockchain.news.
- The source published it on Nov 16, 2026 at 4:17 PM.
Nothing open. Every check passed.
Nvidia pulled the wraps off Rubin HPC on Nov. 16 at SC26, a graphics processor aimed at scientific simulation and data analysis that slots into its Vera Rubin platform. The chip is slated to arrive in 2028, and Nvidia says it will carry more than 50,000 CUDA compute cores plus seventh-generation NVLink for GPU-to-GPU communication.
That NVLink generation is a step up from what's already shipping. Nvidia's existing Vera Rubin lineup — NVL72, a rack-scale box with 72 Rubin GPUs and 36 Vera CPUs, and HGX Rubin NVL8, which links eight Rubin GPUs and pairs with Vera or x86 baseboards — runs on sixth-generation NVLink. Scientific work isn't new territory either: Nvidia also lists Vera Rubin NVL4, four Rubin GPUs and two Vera CPUs built for liquid-cooled MGX servers.
The pitch centers on feeding data to all those cores instead of leaving them idle. Rubin HPC scales data movement alongside compute, Nvidia says, with hardware to accelerate access to scattered data — a design meant to keep cores busy during sparse-matrix operations, sorting and reductions rather than waiting on inputs. Target workloads include climate modeling, engineering simulations, genomic analysis and financial risk. Nvidia's argument: as AI agents churn through more candidate materials, molecules and designs, they'll need more simulations to sort the winners.
What Nvidia didn't say matters too. The announcement skips memory capacity, double-precision arithmetic performance and any application benchmark — so it doesn't establish a like-for-like simulation speedup over an earlier GPU. Also missing: the board, complete system or partner configuration Rubin HPC will ship in, a gap for anyone budgeting around that 2028 window.
Nvidia did tie the chip to CUDA-X, its libraries for accelerated algorithms, and used SC26 to roll out cuPotential for molecular simulations and cuSpa for spatial biology calculations.
TopicsNvidia · Rubin HPC · Vera Rubin · NVLink · CUDA-X · SC26 · NVL72 · HGX Rubin NVL8
Written by AI from verified reporting.

