Why this deserves attention
Intel Arc Pro workstation GPUs—specifically the Arc Pro B60 and B70—can be partitioned into multiple independent virtual functions (VFs) using Single Root I/O Virtualization (SR-IOV) under Proxmox VE without recurring software licensing fees, according to benchmarks and implementation guides published by Level1Techs. The setup relies on custom Linux kernel 6.11 builds and Intel Xe driver patches on Proxmox VE 9.0 to carve single physical cards into hardware-accelerated slices for virtual machines and lightweight containers.
Under this configuration, the Intel Arc Pro B70 (equipped with 32 GB of VRAM at roughly $1,000 per card) supports up to eight virtual functions, allocating approximately 4 GB of dedicated VRAM per instance. The Arc Pro B60 (24 GB VRAM) supports up to seven virtual functions at around 3.4 GB each. By contrast, the older Arc Pro B50 suffered from inconsistent firmware support, officially offering only two virtual functions unless modified with third-party OEM firmware.
In practice, this deployment offers an off-ramp for homelabs and small organizations facing steep licensing overhead from traditional enterprise platforms. Enterprise VDI hardware stacks relying on Nvidia vGPU subscriptions and VMware have reportedly escalated in price from roughly $28,000 to over $117,000. Slicing Intel Arc hardware enables multi-tenant hardware acceleration for 3D modeling, CAD applications like Autodesk Revit, and streaming over RDP or Moonlight. Level1Techs also demonstrated running SR-IOV directly into Kasm Workspaces within Proxmox LXC containers to reduce hypervisor context-switching overhead. During live migration tests on active SR-IOV workloads, Proxmox managed handoffs with roughly 2.5 seconds of display stutter before recovering without crashing the guest operating system.
Despite these gains, the configuration is not yet a turnkey consumer solution. Enabling SR-IOV currently requires manual kernel compilation and out-of-tree driver patching. Furthermore, driver synchronization across guest operating systems has previously shown fragility when versions mismatch, and remote virtualized performance does not match the throughput of a dedicated, high-end bare-metal GPU.