Every Nvidia GPU has 10 to 30 RISC-V cores inside it
3 hours ago
- Nvidia shipped over a billion RISC-V cores in 2024, embedded in GPU management processors rather than shader cores.
- Each Nvidia GPU contains 10 to 40 RISC-V cores that handle tasks like video decoding, power management, and security, replacing the proprietary Falcon microcontrollers.
- The switch from Falcon to RISC-V was driven by the need for 64-bit addressing, data cache, and the ability to run a real operating system.
- Nvidia evaluated Arm, Synopsys ARC, MIPS, and Cadence cores but chose RISC-V for its extensible ISA, security features, and performance-per-area gains (3x performance for 2x area).
- The GPU System Processor (GSP) uses four RV64 cores to run a microkernel (LibOS) that handles driver functions, enabling lower latency and vGPU partitioning.
- Despite RISC-V being open, Nvidia locks down its GPUs with cryptographic signatures, preventing unauthorized firmware execution—frustrating open-source driver developers.
- Open-source drivers (nouveau, Nova) can now control clock speeds thanks to GSP support, a capability lost since Maxwell GPUs in 2014.
- AMD uses a mix of proprietary ISAs (Arm Cortex-A5, Tensilica Xtensa, F32) instead of RISC-V, while Nvidia has secretly integrated RISC-V for over a decade.
- Nvidia announced plans to port CUDA to RISC-V, aiming to let RISC-V CPUs drive Nvidia GPUs, though no timeline was given.