At Hot Chips 2026, Nvidia announced plans to extend CUDA support to RISC-V CPUs. While currently limited to x86-64 and aarch64, this move aims to enable RISC-V processors to drive GPU compute tasks for machine learning applications.

To ensure high performance, Nvidia mandates server-grade hardware meeting RVA23 CPU profiles and specific server SoC specifications. Key requirements include RAS features, security processors, and ACPI support, which was ratified in 2025 via the UEFI forum. This resolves earlier compatibility issues that hindered efficient software porting.

Further technical demands involve PCIe coherency to prevent data caching errors during DMA transfers between CPU and GPU. Nvidia also requires peer-to-peer PCIe communication to avoid performance bottlenecks associated with routing buffers through CPU memory. These strict criteria aim to eliminate "lowest common denominator" inefficiencies.

Additionally, Nvidia outlined requirements for NVLink Fusion, allowing third parties to integrate NVLink C2C links with custom chips. This enables connections between Nvidia GPUs and non-Nvidia CPUs, such as hypothetical RISC-V designs similar to the GB10 architecture, provided they meet all CUDA and framework standards like DOCA and NCCL.