Nvidia announces Rubin HPC, a science GPU with 50,000-plus cores, for 2028
Nvidia unveiled Rubin HPC at the SC26 supercomputing conference, a version of its Rubin GPU rebalanced for FP64 simulation work, due in 2028 as AMD's MI430X targets the same buyers.
Kiln, covering chips, data centers and power
, 2 min read
50,000+CUDA cores in Rubin HPC, about 3x Hopper
What to know
- Nvidia announced Rubin HPC, a GPU in its Vera Rubin platform built for science, engineering and analytics, expected in 2028.
- Nvidia says it has more than 50,000 CUDA cores, supports precisions from FP64 to FP4 and connects over seventh-generation NVLink.
- Nvidia has not disclosed its peak performance, memory capacity or memory bandwidth.
Nvidia is building a GPU for scientific computing again. At the SC26 supercomputing conference in Chicago, Ian Buck, Nvidia's vice president of hyperscale and HPC, announced Rubin HPC, a new GPU in the Vera Rubin platform "built for science, engineering, analytics and AI," according to Nvidia's blog. It is expected in 2028.
ServeTheHome called it Nvidia's first significant HPC-class part since Hopper, which launched almost four years ago. Nvidia's recent data center GPUs have been tuned for AI, which runs on low-precision math, while weather, physics and engineering simulations depend on double-precision (FP64) math.
What Nvidia disclosed
Nvidia says Rubin HPC has more than 50,000 CUDA cores, about three times as many as Hopper. The Register reported the count is nearly twice that of the AI-focused Rubin GPU. The chip supports number formats from FP64 down to FP4, adds acceleration for sparse data, and connects to other GPUs over seventh-generation NVLink, according to Nvidia.
Nvidia has not given peak FP64 performance, memory capacity or memory bandwidth, The Register and ServeTheHome reported. Neither outlet listed any customers or planned systems. ServeTheHome expects HBM4E memory based on Nvidia's rendering, but Nvidia has not confirmed that.
Dan Ernst, Nvidia's senior director of supercomputing, told The Register the chip rebalances the architecture for simulation, engineering and high-performance analytics. The Register estimated from Nvidia's description that Rubin HPC's FP64 vector performance would be at least 86% higher than the AI Rubin's, which it put at 33 teraflops. That is the outlet's calculation, not an Nvidia figure. Nvidia has instead been promoting FP64 emulation on its AI chips, which it says lets Rubin reach up to 200 emulated FP64 matrix teraflops; Ernst said HPC hardware should still offer native FP64 matrix performance, according to The Register.
The AMD comparison
AMD is aiming at the same buyers sooner. Its Instinct MI430X, which AMD described in a blog post the week before SC26, carries 432 GB of HBM4 memory, 23 TB/s of memory bandwidth and 288 teraflops of peak FP64, according to The Register. AMD builds separate AI and HPC variants from chiplets rather than relying on emulation. The MI430X is going into EuroHPC's Alice Recoque system and Oak Ridge National Laboratory's upcoming Discovery supercomputer, The Register reported. ServeTheHome said the MI430X is due in 2027, a year ahead of Rubin HPC.
Other SC26 announcements
Nvidia also expanded its Vera CPU platform for HPC, and said the German Climate Computing Center (DKRZ) chose Vera CPUs for its HLRE-5 system, expected in early 2028. It announced Nemotron for Science, an open model and chemistry dataset, and a quantum computing collaboration with IBM and Fermilab, according to its blog.
Why it matters
National labs and research centers buy supercomputers years ahead, and many of their codes need native FP64 performance that AI-focused GPUs have cut back. Rubin HPC is Nvidia's answer to AMD's lead in that market, but buyers planning 2027 and 2028 systems still lack its key specifications. For background on how GPUs differ from CPUs, see /stories/gpu-vs-cpu-explained, and on the memory these chips depend on, /stories/hbm-memory-explained. Nvidia's latest filings show $390 billion in supply commitments, and AMD's MI455X is behind TensorWave's 457 MW deal for OpenAI.
Sources
- Ian Buck Kicks Off NVIDIA's Week at SC26 as Science Enters the Age of Agents, NVIDIA Blog, November 17, 2026
- GPU giants throw HPC users some tasty crumbs, The Register, November 17, 2026
- NVIDIA Announces Rubin HPC: Science-Focused Server GPU to Arrive in 2028, ServeTheHome, November 17, 2026
NvidiaAI chips and data centers
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.