Nvidia intends to port Cuda to RISC-V, but this is less revolutionary than it sounds. Speaking at the Fifth RISC-V China Summit, Nvidia VP Frans Sijstermans stated the company is cooperating with hardware partners to support the open architecture in the standard Cuda version. However, this doesn’t imply Nvidia will support RISC-V chips as alternatives to its GPUs.
What Does Cuda for RISC-V Mean?
Cuda is how coders target Nvidia GPUs, leading a few observers to erroneously conclude that adding RISC-V support means replacing GPU devices as the target. Cuda is a framework, not a programming language. Developers program in a conventional language (e.g., C++ or Python) extended for parallelism and supplemented by libraries providing common functions. Cuda compiles the software, dividing object code between a CPU host and a GPU device and coordinating their joint execution.
The CPU host is where Nvidia intends to support RISC-V, placing the architecture alongside Arm and x86. Although companies such as Tenstorrent and Meta employ RISC-V cores as central building blocks in their NPUs, Nvidia will not target such RISC-V devices. (Blackwell and other Nvidia GPUs integrate tens of invisible small RISC-V cores, but these handle low-level functions and aren’t relevant to this Cuda discussion.)
Targeting RISC-V is a large undertaking. Drivers, development and build tools, and numerous libraries must all be ported. Moreover, few standard RISC-V processor boards are available. Nvidia is developing Cuda on SiFive’s HiFive Premier P550 board, and we expect it to shift to a platform implementing the RVA23 profile when one becomes available. Ratified last year, RVA23 defines a 64-bit architecture for application processing and mandates vector and hypervisor extensions.
Underscoring the industry coalescing around RVA23, Canonical stated that the profile will be Ubuntu’s required baseline starting with the Linux distribution’s 25.10 version. Meanwhile, Red Hat is previewing its Enterprise Linux 10, initially targeting the HiFive Premier P550 board. Despite rapid progress, the RISC-V ecosystem remains immature, and application software and middleware are lacking.
Why Support RISC-V?
Nvidia has adopted Arm for its Grace processor and employs Intel Xeon in some systems. No one supplies RISC-V servers, and no suitable RISC-V processors are in production. However, we expect this to change. As China decouples from the West, companies are likely to develop RISC-V server processors. Several outside of China have developed high-performance RISC-V cores, but they have yet to be deployed in commercial silicon. As the RISC-V ecosystem matures, companies worldwide may switch to it.
Bottom Line
By backing RISC-V, Nvidia boosts the architecture and positions its GPUs as the foremost graphics, HPC, and AI accelerators for mavericks pursuing the Arm and x86 alternative. As much development work as Nvidia has ahead of it, RISC-V maturity will be the factor limiting Cuda’s support for the architecture.