
KEY POINTS
- Nvidia unveiled Rubin HPC at Supercomputing 2026, with more than 50,000 CUDA cores and an expected 2028 arrival.
- Nvidia has not disclosed peak performance, memory capacity or bandwidth for the new chip.
- AMD's Instinct MI430X offers 432 GB of HBM4, 23 TB/s of bandwidth and 288 teraFLOPS of peak double precision.
Nvidia has announced a version of its Rubin GPU built for high performance computing (HPC), presenting it at the Supercomputing 2026 conference this week. The Register reported that the chip is due sometime in 2028 and is meant to serve simulation, engineering and scientific analytics users.
According to the report, Rubin HPC will have more than 50,000 CUDA cores, nearly double the count of the AI-focused Rubin that Nvidia revealed earlier this year. The Register's own calculation puts the figure at about 53,248, which would mean at least 86 percent higher double-precision vector performance.
Dan Ernst, Nvidia's senior director of supercomputing, described the design as a rebalanced architecture for workloads common in scientific computing. He said Nvidia builds variants of a platform to target specific sub-domains, and that Rubin HPC follows the same approach Nvidia took last year when it tuned its B300 accelerator for AI inference.
Nvidia has shared few details. It has not disclosed peak performance, memory capacity or memory bandwidth for the new part. Ernst also gave no specifics, but said the HPC hardware should deliver native FP64 matrix performance, so customers would not depend only on emulation.
The concern among HPC users has a clear background. Since Blackwell Ultra, Nvidia has favored lower-precision throughput over double-precision matrix acceleration. That chip reached 15 petaFLOPS at FP4 but only 1.3 teraFLOPS at FP64. The standard Rubin delivers 33 teraFLOPS at FP64, below both Blackwell and the four-year-old Hopper generation.
For dense matrix workloads, such as those represented by the High-Performance Linpack benchmark, Nvidia has turned to FP64 emulation based on the Ozaki Scheme. The method splits FP64 matrix operations into a series of lower-precision operations on tensor cores. Nvidia says Rubin can reach up to 200 emulated FP64 matrix teraFLOPS. The company also argues that the higher ratio of memory bandwidth to compute should help sparse workloads measured by the HPCG benchmark.
AMD is pursuing the same market. In a blog post published late last week, it detailed the Instinct MI430X, an HPC-tuned member of the MI400 series. The chip carries 432 GB of HBM4 memory, 23 TB/s of memory bandwidth and 288 teraFLOPS of peak double-precision performance.
AMD does not rely on emulation. It used its modular chiplet design to produce two versions of the MI400 line, one for AI and one for scientific computing. The company has also projected results on several Top500 benchmarks, saying systems such as EuroHPC's Alice Recoque and Oak Ridge National Laboratory's upcoming Discovery would gain significantly over setups built on its MI300A and MI250X chips.