Nvidia announces at the RISC-V summit in China that CUDA support for the architecture is coming.
Nvidia is bringing CUDA to RISC-V. The company made this announcement at the RISC-V Summit in China. The announcement being made in China is not surprising, as the country is increasingly building its own chips in an effort to become independent from Western models, and the open-source RISC-V architecture is well-suited for this purpose.
CUDA and RISC-V
Nvidia CUDA is the abstraction layer between Nvidia hardware on one side and applications on the other. The CUDA ecosystem can be seen as the secret sauce that contributes to the dominant position of the GPU manufacturer today. Around CUDA, an ecosystem has emerged that is so pervasive that manufacturers like AMD cannot compete with powerful hardware alone.
read also
Nvidia to Support CUDA on RISC-V
RISC-V, on the other hand, is an alternative CPU architecture. The most significant difference from x86 and ARM is not functional but relates to openness. RISC-V is open source. Those who build RISC-V chips owe no licensing fees to anyone.
Architecture-agnostic
Nvidia initially built CUDA for x86 but also supports ARM. The manufacturer’s Grace Hopper superchip, for example, combines ARM cores with its own GPU. ARM-based servers from other manufacturers can also work with Nvidia GPUs and the CUDA software stack.
This compatibility is now coming to RISC-V, implying that manufacturers can build servers with a RISC-V CPU, linked to Nvidia GPUs, running a CUDA-based stack just like with x86 and ARM. The support is essential for the relevance of RISC-V in an AI context.
Several Chinese companies, including Alibaba, are actively working on RISC-V-based CPU cores. Nvidia recently received renewed permission to sell (restricted) GPUs to China. The company wants to capitalize on this, and compatibility with homegrown Chinese processors should not be a barrier. When exactly the RISC-V support will be implemented is unclear.