Nvidia announces Rubin HPC GPU for 2028 as AMD details MI430X
At Supercomputing 2026, Nvidia revealed an HPC-focused Rubin variant with more than 50,000 CUDA cores, while AMD disclosed an Instinct MI430X rated at 288 teraFLOPS of double precision.
· 2 min read

- Nvidia unveiled Rubin HPC at Supercomputing 2026, with more than 50,000 CUDA cores and an expected arrival in 2028.
- 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 FP64.
Nvidia has introduced a version of its Rubin GPU built for high performance computing (HPC), announced at Supercomputing 2026 this week. The Register reported that the chip has more than 50,000 CUDA cores, nearly double the count of the AI-oriented Rubin that Nvidia presented earlier this year, and that it is due sometime in 2028.
Dan Ernst, Nvidia's senior director of supercomputing, said the chip uses a rebalanced architecture tuned for simulation, engineering and high-performance analytics, the workloads common in scientific computing. He compared it to last year's B300 accelerator, which Nvidia optimized for AI inferencing. "Rubin HPC is the same idea," Ernst told the outlet, adding that Nvidia is building variants within the Rubin platform for specific sub-domains of applications.
Few specifics are available. According to The Register, Nvidia has not published peak performance figures, memory capacity or memory bandwidth for the new part. The outlet estimates the core count is closer to 53,248 and says that would mean at least 86 percent higher double-precision vector performance.
The announcement follows concerns in the HPC community. Starting with Blackwell Ultra, the B300 and GB300, Nvidia gave up double-precision matrix acceleration in favor of higher throughput at lower precision. That chip reached 15 petaFLOPS at FP4 but only 1.3 teraFLOPS at FP64.
The standard Rubin continued the pattern, with 33 teraFLOPS at FP64, below both Blackwell and the four-year-old Hopper accelerators. Nvidia argued that sparse, vector-heavy workloads, as measured by the HPCG benchmark, should still gain from a higher ratio of memory bandwidth to compute. For dense matrix workloads such as those represented by HPL, it turned to FP64 emulation based on an algorithm derived from the Ozaki Scheme. Nvidia says Rubin can reach up to 200 emulated FP64 matrix teraFLOPS.
Ernst gave no details, but told the outlet that Nvidia's HPC hardware should deliver native FP64 matrix performance, so customers would not need to depend entirely on emulation.
AMD is also preparing an HPC-specific part. In a blog post published late last week, it described the Instinct MI430X, a variant of its MI400 series with 432 GB of HBM4, 23 TB/s of memory bandwidth and 288 teraFLOPS of peak double-precision performance. AMD does not use FP64 emulation, and instead used its modular chiplet design to create separate AI and HPC versions of the MI400 line.
AMD has also shared projected results for several headline Top500 benchmarks. It says systems such as EuroHPC's Alice Recoque and Oak Ridge National Laboratory's upcoming Discovery would see significant gains by moving up from MI300A and MI250X-based hardware, according to The Register.