GPU Para-Virtualization on Xen

on YouTube ↗33m

" summary

Updated Xen GPU stack moves from Virgl/Venus to a virtio native context that forwards native DRM calls so guest RadeonSi (OpenGL) and Radv (Vulkan) run directly on AMDGPU hardware, reducing Mesa-in-dom0 overhead. Prototype adds PVH dom0 passthrough, blob memory via QEMU, and a ROCm/OpenCL prototype for compute. Benchmarks show many cases at 80–90% of bare metal and gaming FPS improves ~10–20% versus Venus. Design includes SVM/userptr plumbing, MMIO/virt MMBAR direct mapping work, and ongoing upstream patches to complete dynamic mapping and security hardening.

AI-written summary · about

" chapters (6)

1. 0:00Introduction and roadmap
2. 6:38Graphics virtualization
3. 14:01Compute virtualization
4. 14:29Native context and RoCm
5. 23:21Future steps and hardware
6. 25:08Demo and Q&A

" related talks

• 36mCurrent State of Graphics Virtualization UpstreamDaniel Stone · Automotive Linux Summit @ Open Source Summit Japan 2022
• 22mKernel-Native Packet Processing on AMD GPUs: XDP Offload without Userspace RuntimesTaehee Yoo · Linux Plumbers Conference 2025
• 20mPassthrough/Headless GPU Gets AheadTina Zhang, Vivek Kasireddy · KVM Forum 2021
• 29mXenGT: A software based Intel Graphics Virtualization SolutionHaitao Shan · Xen Project User Summit 2013
• 08mNova GPU Workshop: vGPU + VFIO: status and 2026 plansZhi Wang · Linux Plumbers Conference 2025
• 22mAMD-vIOMMU: A Hardware-assisted Virtual IOMMU TechnologyWei Huang, Suravee Suthikulpanit · KVM Forum 2020
• 26mAnalysis of AMD HW-assisted vIOMMU Implementation and PerformanceWei Huang, Suravee Suthikulpanit · KVM Forum 2021
• 20mSpeed Up Creation of a VM With Pass Through GPULiang Li · KVM Forum 2020