Nvidia Unveils Rubin HPC with Over 50,000 CUDA Cores

Nvidia Corporation introduced the new Rubin HPC graphics accelerator at the Supercomputing 2026 conference, designed for high-performance computing (HPC). The device will feature over 50,000 CUDA cores and is aimed at scientific modeling, engineering calculations, and large-scale data analysis. Its shipments are expected to begin in 2028.
Features of Nvidia Rubin HPC
Rubin HPC will be part of the Nvidia Vera Rubin computing platform. According to Dan Ernst, Nvidia's senior director of supercomputing products, the architecture of the new accelerator is optimized for tasks typical of scientific organizations and research centers.
One of the main differences of Rubin HPC will be the increased number of computing cores. Nvidia confirmed that the accelerator will have over 50,000 CUDA cores—almost twice as many as the previously introduced Rubin version, which was primarily focused on artificial intelligence tasks.
Additionally, developers have provided special mechanisms to accelerate the processing of sparse data. These are designed for more efficient data transfer between memory and computing units, which is especially important for scientific calculations involving large data volumes.
The seventh-generation NVLink interface will be used for communication between graphics accelerators.
However, Nvidia has not yet disclosed the exact performance of Rubin HPC in double-precision FP64 calculations, the amount of video memory, or its bandwidth.
Why FP64 performance is important for scientific computing
In recent years, Nvidia has focused significantly on accelerating reduced-precision operations, which are in demand for training and operating artificial intelligence models. However, scientific modeling and engineering calculations often require high numerical precision, which is why FP64 calculations remain important for them.
The previously introduced Rubin has a declared vector FP64 performance of 33 TFLOPS. This is lower than the corresponding figures for Nvidia Blackwell and Hopper.
For some matrix operations, Nvidia uses a different approach—emulating FP64 using algorithms based on the Ozaki scheme. This method allows distributing computations among faster data processing units and achieving high performance in certain tasks.
For the standard Rubin, the company indicates up to 200 TFLOPS in FP64 matrix operations using such emulation.
However, this figure does not apply to the new Rubin HPC. Nvidia has not yet presented its exact specifications.
Additional CUDA cores may increase the computational performance of the new accelerator, but the actual gain will depend on the architecture, frequencies, memory performance, and the specifics of particular applications.
AMD reveals Instinct MI430X specifications for HPC
Simultaneously, AMD has provided additional details about the Instinct MI430X graphics accelerator, which is also intended for high-performance and scientific computing.
The device is built on the CDNA 5 architecture and is designed to perform complex computational tasks, including physical process modeling and processing large scientific data sets.
According to the specifications published by AMD, the Instinct MI430X will feature 432 GB of HBM4 memory with a bandwidth of up to 23.3 TB/s. The declared peak native FP64 performance reaches 288 TFLOPS.
Unlike software emulation, this figure refers to double-precision calculations performed by the accelerator's hardware.
AMD also presented preliminary performance estimates for the MI430X in scientific tests, including HPL and HPCG. These figures are based on the company's calculations and will require confirmation once serial devices become available.
The AMD Instinct MI430X is expected to be released in 2027, while Nvidia plans to make Rubin HPC available in 2028.
Both manufacturers are developing specialized accelerators for scientific computing, but direct performance comparison is currently limited. AMD has disclosed key specifications for the MI430X, while Nvidia has yet to present detailed specifications for Rubin HPC.
This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.




