AMD Medusa-era Ryzen AI NPUs may support SR-IOV
AMD is preparing SR-IOV support for its next-generation Ryzen AI NPUs. According to Phoronix, a new Linux patch series adds Single Root I/O Virtualization Virtual Function support for AMD’s upcoming AIE4 platform through the AMDXDNA accelerator driver.
SR-IOV is a PCI Express virtualization feature. It allows one physical device to expose multiple virtual functions, which can then be assigned to guest virtual machines. For an NPU, this could let a VM access the Ryzen AI accelerator more directly for local AI workloads, model testing, or software validation.
This is different from standard PCIe passthrough. Passthrough usually assigns the whole device to one VM. SR-IOV can split a supported device into several virtual functions, depending on hardware, firmware, host driver, and guest driver support. Needless to say, it’s a very good feature for graphics cards, as long as it is supported.
Source: Kernel.org
XDNA3 or 4?
AMD has not confirmed the retail name for AIE4. However, this looks like a likely candidate for XDNA 3, the expected successor to the XDNA 2 NPU used by current Ryzen AI processors. The timing also fits Medusa, AMD’s expected Zen 6 client APU generation, which follows Gorgon Point and Ryzen AI 400 (also using XDNA2).
GPU SR-IOV support
When it comes to GPUs, support is in a different state. AMD supports SR-IOV through its MxGPU stack, but official support is aimed at Instinct accelerators and selected Radeon PRO virtualization products. Consumer Radeon RX cards can often be passed through to a VM as a whole device, but AMD does not officially support splitting them across multiple VMs through SR-IOV.
If memory serves me well, GeForce RTX does not support SR-IOV, only data-center and RTX PRO series. Intel does not support SR-IOV with Arc GPUs, only with Arc Pro B-series and its Data Center series.
