Hot Chips 2026: CUDA Targets RISC-V

CUDA, the most important software framework in GPU computing and machine learning, currently supports only x86-64 and aarch64 CPUs. Nvidia is preparing to extend this support to RISC-V CPUs, presenting clear requirements for server-grade CPU and platform compliance.

Nvidia primarily requires compliance with the RVA23 CPU profile and RISC-V server SoC and platform specifications. These specifications include elements essential for server-grade hardware, such as Reliability, Availability, and Serviceability (RAS) features and dedicated security processors.

Beyond existing specifications, Nvidia requires additional hardware features for the efficient operation of CUDA software. In particular, it cited predication support in vector extensions as an example, to avoid the 'least common denominator' problem where inefficient code is deployed because performance-boosting extensions cannot be guaranteed.

Additionally, ACPI support is essential for hardware feature detection and power, performance, and thermal management. The absence of ACPI on RISC-V hardware was previously an issue, but this has been resolved with the addition of RISC-V ACPI support by the UEFI Forum in 2025 and the ratification of the RISC-V BRS (Boot and Runtime Services) specification.

Finally, PCIe coherency is required. This is to prevent issues where DMA engines read stale data from DRAM when transferring data stored in CPU caches to the GPU, or where CPUs read stale data from caches when retrieving results from the GPU. Nvidia aims to avoid these complex software-level cache invalidation processes by leveraging PCIe coherency, a standard feature of server CPUs.