UK — The Register
GPU giants throw HPC users some tasty crumbs
Chips1 min read
Nvidia has shown a new version of its Rubin GPU chip made for high performance computing (HPC), the field of very fast science and engineering computers. It is meant to compete with AMD's new HPC chips.
Nvidia's new HPC chip
At the Supercomputing 2026 event, Nvidia unveiled Rubin HPC, a version of its Rubin GPU (graphics processor, a chip that does many calculations at once) built for science and engineering rather than AI. Nvidia says it will arrive sometime in 2028.
Key numbers and facts
- Rubin HPC promises more than 50,000 CUDA cores (small units that do the math), nearly twice as many as the AI-focused Rubin.
- Nvidia says Rubin can reach up to 200 emulated FP64 teraFLOPS for certain matrix work. FP64 is a high-precision number format that scientists need.
- Based on core count, the article's writer estimates the chip should deliver at least 86 percent higher vector performance at double precision.
- The earlier Blackwell Ultra (B300) chip reached 15 petaFLOPS at FP4 (a low-precision format used for AI) but only 1.3 teraFLOPS at FP64.
Why scientists were worried
Recent Nvidia chips focused on AI and gave much less power for high-precision math. Scientists worried Nvidia might ignore their needs. Nvidia now uses a technique called FP64 emulation, which uses cheaper calculations to act like high-precision math and can run faster on some tasks.
Why it matters
The chip shows that Nvidia still cares about scientific computing, not only AI. Details are still thin, though, so the real speed and power will only be known later. AMD is also building HPC chips, so competition in this market is growing.
What to watch
- Nvidia has not yet shared peak performance, memory capacity, or memory bandwidth for Rubin HPC.
- Whether Nvidia's HPC chip has native FP64 performance, not just emulated performance, as Nvidia's engineers suggested.
- How AMD's HPC-tuned MI400-series GPUs compare, including AMD's MI430X with 288 teraFLOPS.
- The actual launch in 2028.
Glossary
- Nvidia
- A company that makes GPU chips used for AI, gaming, and scientific computing.
- Rubin
- Nvidia's GPU chip family; this story covers a version made for HPC.
- HPC
- High performance computing: very fast computers used for science, engineering, and simulations.
- FP64
- A high-precision number format that scientists need for accurate calculations.
- CUDA cores
- Small processing units inside Nvidia GPUs that do many calculations at the same time.
- AMD
- A chip company that makes CPUs and GPUs, including its MI400-series for HPC.