INFRA Signal 584 2 feeds carried it
Nvidia Targets CUDA for RISC-V CPUs With Server-Grade Hardware Requirements
↗
Nvidia is exploring CUDA support for RISC-V CPUs, but only if the hardware meets strict server-grade requirements including RVA23 compliance, ACPI, PCIe coherency, and peer-to-peer PCIe communication.
This could open GPU compute workloads to RISC-V server platforms, but Nvidia's requirements exclude virtually all existing RISC-V consumer hardware and demand features that avoid lowest-common-denominator code generation. The specifications set a concrete bar for any RISC-V vendor aiming to pair their CPUs with Nvidia GPUs.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Nvidia requires RVA23 CPU profile, RISC-V server SoC and platform specifications, RAS features, and a specialized security processor for CUDA on RISC-V.
PCIe coherency and peer-to-peer PCIe communication are mandatory to avoid cache consistency bugs and performance penalties in the CUDA stack.
NVLink Fusion imposes all CUDA requirements plus DOCA and NCCL framework support and a close partnership with Nvidia.
THE CLUSTER