Nvidia RISC-V and CUDA ecosystem expansion

Nvidia is expanding its software and hardware ecosystem by integrating support for the RISC-V architecture. To ensure efficient performance and avoid a “lowest common denominator” approach, the company has established specific requirements for server-grade CPUs, including adherence to RVA23 standards, PCIe coherency, and reliability, availability, and serviceability (RAS) features.

This integration aims to position RISC-V as a third architecture option for accelerated servers, alongside x86 and Arm. Nvidia’s technical roadmap involves splitting workloads so that compute-intensive functions are parallelized on the GPU while the host CPU manages sequential code and system services. Key technical components for this integration include hardware-managed PCIe coherence and the use of NVLink Fusion to connect custom CPUs and accelerators through a unified memory architecture.

Parallel to these hardware developments, Nvidia is enhancing its CUDA-X suite of GPU-accelerated libraries. Recent updates include new solver libraries for quantum chemistry and significant performance gains in computational engineering and lithography. Additionally, the company’s cuDF library has been integrated into Cloudera to accelerate Apache Spark workloads.