Let's be clear about why this matters: CUDA is the single biggest moat in compute. It's been x86-64 and AArch64 only, forever. RISC-V has been the open-source underdog promising architectural freedom. Now NVIDIA is signalling they'll let CUDA cross the moat.
At Hot Chips 2026, NVIDIA laid out exactly what a RISC-V CPU needs to run CUDA. The requirements are blunt:
- RVA23 profile compliance — the latest RISC-V application-class baseline
- ACPI — for hardware discovery and power/thermal management. NVIDIA's team was candid: they weren't happy RISC-V didn't have ACPI when they started the port
- PCIe coherency — to solve memory ordering problems where CPU cache has data DMA engines can't see
- Peer-to-peer PCIe — so buffers between devices don't have to bounce through CPU memory
Missing any of these means CUDA won't perform. NVIDIA explicitly said they won't support a "lowest common denominator" — they want server-grade or nothing.
The broader play here is NVLink Fusion, which lets other chipmakers implement NVIDIA's NVLink IP on their own silicon. SiFive is the named partner, with plans to demo a system running CUDA at the show.
The Verdict
This isn't about your RISC-V SBC getting GPU compute next year. It's about the data centre. If RISC-V servers start shipping with CUDA support, the CPU architecture lock-in that's held since the 90s starts cracking. The wall around NVIDIA's garden just got a door — and the key is compliance with a two-page requirements document.
Watch SiFive's demo closely. If it works, the CPU landscape just got interesting again.