Hot Chips 2026: CUDA Targets RISC-V – By Chester Lam (chipsandcheese.com)

Nvidia has announced plans to extend CUDA support to RISC-V CPUs, aiming to boost GPU compute capabilities in this emerging architecture. Historically, CUDA has only supported x86-64 and aarch64 platforms, making this a significant expansion. To effectively integrate RISC-V CPUs with CUDA, Nvidia has set stringent requirements, including the need for a RVA23 CPU that adheres to specific server SoC and platform specifications like RAS features, specialized security processors, and PCIe coherency. These requirements are crucial as they ensure efficient data handling between CPUs and GPUs, thereby avoiding performance degradation due to stale caches or inadequate data synchronization. This development comes with the recognition that RISC-V's software ecosystem still lags behind its competitors, making widespread implementation a challenge. Nvidia's emphasis on server-grade prerequisites suggests that initial RISC-V systems capable of running CUDA will likely cater to high-performance applications rather than consumer-grade setups. Although Nvidia admitted that existing consumer hardware may not meet these specifications in the near term, it hinted at partnerships, such as with SiFive, to demonstrate working systems. This initiative underscores Nvidia's commitment to enhancing GPU performance across varied architectures while also highlighting the potential need for more flexible requirements to enable broader accessibility of CUDA within the RISC-V landscape.