Nvidia is pushing back against the perception that it is moving away from high performance computing and 64-bit precision, clarifying that recent product choices do not signal an exit from the space. The company told HPCWire that 64-bit floating point data remains central to its roadmap, even as recent architectures such as Hopper and Blackwell have emphasized lower precision formats more aligned with acceleration for artificial intelligence workloads.
Dion Harris, senior director of high performance computing and artificial intelligence hyperscale infrastructure solutions at Nvidia, said the company is “definitely looking to bring some additional [FP64] capabilities in our future gen architectures” and stressed that Nvidia is “very serious about making sure that we can deliver the required performance to power those simulation workloads.” The comments are aimed at users who rely on sustained double precision throughput, particularly in scientific and engineering domains, and who have been concerned by stagnating fp64 metrics in newer flagship accelerators.
The acceleration of 64-bit floating-point data paths is described as crucial for the high performance computing community, with the life sciences called out as a key beneficiary. Users have noted that when a workload demands sustained high-precision support, Nvidia’s recent generations have not met expectations. For comparison, Nvidia’s current most powerful B300 “Blackwell Ultra” accelerator achieves only 1.2 TeraFLOPS of FP64 performance. In contrast, the older H200 “Hopper” reaches an impressive 34 TeraFLOPS of FP64 compute at its peak. For FP8 low-precision, the B300 delivers 9 PetaFLOPS, while the H200 provides 3.958 PetaFLOPS. These figures highlight how Nvidia has so far optimized its newest platforms for lower precision formats, even as it now publicly commits to improving double precision capabilities in its next generation designs.
