AMD Details Radeon GPU Passthrough for Proxmox AI Workloads

An Oct. 19 AMD guide outlines how to assign Radeon GPUs to Proxmox VE virtual machines running ROCm and PyTorch. The workflow supports isolated development, repeatable continuous integration jobs and environments that operate without a network connection.

The setup gives one virtual machine exclusive use of a physical GPU. While the device is assigned, the host cannot use it; the guest operating system loads ROCm for the GPU. The instructions cover integrated Radeon graphics in Ryzen AI Max systems as well as discrete Radeon AI PRO cards.

AMD’s examples use Proxmox VE 9.1 in two configurations: a Ryzen AI Max+ 395 system with Radeon 8060S integrated graphics, and an AMD Ryzen Threadripper PRO 9965WX host with four Radeon AI PRO R9700 GPUs. Both use Ubuntu 24.04 HWE, ROCm 7.2 and PyTorch 2.9.1+rocm7.2.

The virtual machine layer provides templates, snapshots, cloning and REST APIs for managing GPU workloads. Those tools can support repeatable CI runs, restore a machine to an earlier state or keep tenants separate. Bare-metal systems may suit users who prioritize maximum performance or need the host to access the GPU directly. Containers can be simpler for some Linux workloads, while offering less GPU fault isolation and no equivalent driver-state reset in this comparison.

Platform memory ceilings are separate from the reported test configurations. Ryzen AI Max+ 395 systems support up to 128 GB of unified LPDDR5X memory; the Ryzen AI Max+ PRO 495 ceiling is 192 GB.

The procedure covers BIOS settings, IOMMU enablement, VFIO device binding, virtual machine creation and ROCm installation. For integrated graphics, users must stop the VM after installing ROCm, reboot the Proxmox host, then start the VM again. With a discrete GPU, restarting the guest VM is sufficient.