Particle.news

NVIDIA Unveils Rubin HPC GPU With More Than 50,000 CUDA Cores

The company says the design combines sparse-data acceleration and seventh-generation NVLink to keep many cores fed and speed scientific simulation with availability expected in 2028.

Overview

  • NVIDIA announced Rubin HPC at SC26 as a new GPU in its Vera Rubin platform built to accelerate simulation, data processing and agent-driven discovery in science and engineering.
  • NVIDIA says a single Rubin HPC GPU will carry a record core count of more than 50,000 CUDA cores and includes sparse-data acceleration to reduce idle core time.
  • The company reported seventh-generation NVLink for faster GPU-to-GPU connectivity and described balanced design choices that link many cores with memory and I/O to improve simulation throughput.
  • Independent reporting analyzed NVIDIA images and suggested system-level designs such as multi-chip Superchips, NVLink switches and trays that could scale core counts into the hundreds of thousands, but those system details are speculative.
  • NVIDIA lists Rubin HPC availability as expected in 2028, and analysts note delivery risk based on the long lead time and NVIDIA’s prior Rubin CPX cancellation, while CUDA-X libraries are offered to speed real-world adoption.