GPU giants throw HPC users some tasty crumbs
First reported by The Register ·
Nvidia now offers an HPC-specific GPU that aims to restore high double-precision performance, which was previously reduced for AI workloads.
Nvidia has announced a new variant of its Rubin GPUs specifically for high-performance computing (HPC) workloads, dubbed Rubin HPC. Unveiled at Supercomputing 2026, this specialized chip will feature over 50,000 CUDA cores, nearly double the AI-focused version, and is expected to arrive in 2028. Nvidia claims the Rubin HPC's rebalanced architecture is optimized for scientific computing, simulation, and analytics. This move addresses concerns that Nvidia might deprioritize HPC after shifting its focus to AI, which led to a decrease in double-precision (FP64) performance in previous chip generations like Blackwell Ultra. The company stated that while it previously reduced FP64 capabilities to boost AI throughput, the Rubin HPC will offer native FP64 matrix performance in addition to enhanced vector performance. AMD is also releasing its own HPC-tuned GPU, the Instinct MI430X, boasting significant double-precision performance without relying on FP64 emulation.
Nvidia's introduction of a specialized Rubin HPC GPU signifies a strategic effort to cater to the distinct needs of the scientific computing and HPC communities, which often require superior double-precision (FP64) performance that differs from AI-centric workloads. The company's previous focus on AI throughput had led to a decrease in FP64 capabilities, a move that concerned HPC users. By rebalancing the architecture and promising native FP64 performance, Nvidia aims to retain its standing in a market segment that was beginning to look towards alternatives like AMD's MI430X, which emphasizes raw FP64 power.
This dual-pronged approach, offering both AI-optimized and HPC-optimized variants within the same GPU platform, allows Nvidia to maximize its market reach without sacrificing performance in critical scientific applications. For users, this means a renewed focus on hardware tailored to their specific computational demands, potentially leading to significant performance gains in simulation, modeling, and complex data analysis, and preventing a scenario where scientific computing is sidelined in favor of the AI boom.
AI-written summary. May contain errors.