GPU Limitations in WSL2¶

GPU VRAM is shared between Windows and WSL2. There is no partitioning or reservation mechanism.

Windows desktop compositing (DWM): ~200-500 MB
Effectively ~15.5 GB available for CUDA workloads with the 16 GB RTX 4060
.wslconfig memory setting controls system RAM, not VRAM

nvidia-smi Limitations¶

nvidia-smi in WSL2 returns incomplete data vs bare metal:

Metric	Available?
GPU name, driver version	Yes
Temperature, power draw	Yes
VRAM total/used/free	Yes
GPU utilization %	No
Active process list	No
Clock speeds	Partial

For monitoring, use nvitop or gpustat with caveats about missing utilization data.

Vulkan¶

No native NVIDIA Vulkan in WSL2.
Vulkan is available only through Mesa's experimental Dozen driver (D3D12-to-Vulkan translation layer). Not conformant, testing-only.
Dozen driver does not ship in standard distro packages; requires rebuilding Mesa from source with -D vulkan-drivers=microsoft-experimental.
No official timeline from Microsoft for proper Vulkan support.
Impact: Vulkan-based renderers and some game engines won't work. CUDA compute is unaffected.

OpenGL¶

OpenGL is available through WSLg (Wayland/X11 forwarding via RDP).
Not GPU-accelerated for display rendering -- uses software rendering (llvmpipe) or D3D12 Mesa backend.
OpenGL-CUDA interop is not supported.
Impact: negligible for headless compute. Affects only GUI OpenGL applications.

Display Output¶

WSL2 has no direct display output from the GPU.
GUI apps render through WSLg (remote desktop protocol to Windows desktop).
For headless compute (your use case), this is irrelevant.

CUDA Feature Gaps¶

Feature	Status
Unified Memory (full managed)	Not available
Concurrent CPU/GPU access to managed memory	Not supported
Pinned system memory	Restricted availability
CUDA IPC (legacy APIs)	Requires driver R510+
OpenGL-CUDA interop	Not supported
Multi-GPU device selection in containers	Only `--gpus all`

Performance Overhead¶

Compute-bound: 5-10% overhead (long GPU kernels, matrix ops)
Kernel-launch-heavy: Up to 33% overhead (many small GPU kernels)
I/O-bound: Higher overhead due to WSL2 filesystem virtualization
Recommendation: Store data on ext4 (/home/...), not NTFS (/mnt/c/...)

NVML and Monitoring Tools¶

The NVIDIA Management Library (NVML) has reduced functionality in WSL2. Tools that depend on NVML (like some Prometheus exporters) may not report all metrics. nvidia-smi dmon and nvidia-smi pmon have limited output.

Windows Driver Updates¶

The GPU driver exists only on the Windows side. WSL2 automatically gets updated stubs when the Windows driver is updated. You don't need to (and must not) update any driver inside WSL2.

If GPU access breaks after a Windows driver update, wsl --shutdown and restart usually fixes it.