CUDA Toolkit 11.8
supported- Released
- 2022-10
- Minimum driver — Linux
- 520.61.05
- Minimum driver — Windows
- 522.06
- Supported GPU architectures
- Compute capability (sm_XX)
- Commercial GPU names
- Compatible cuDNN
WSL2 & Docker
WSL2: this version's own driver minimum above already clears the R495 driver branch floor NVIDIA's CUDA-on-WSL2 guide requires for GPU support under WSL2 (docs.nvidia.com/cuda/wsl-user-guide), so no extra driver bump is needed beyond what's listed above -- just remember never to install a separate Linux driver inside WSL2 itself (the Windows driver is exposed into WSL2 automatically). Docker: a host driver >= the Linux minimum above plus the NVIDIA Container Toolkit is sufficient to run this CUDA version in a container -- no extra Docker-specific driver floor beyond that number.
Architecture support corrected 2026-09 against NVIDIA's archived release notes (10.2, 11.0, 12.0, 13.0): Kepler sm_30/sm_32 were dropped in CUDA 11.0 (deprecated since 10.2), Kepler sm_35/sm_37 were removed in 12.0, and Maxwell/Pascal/Volta were removed in 13.0 (still supported, as deprecated, throughout 12.x). Windows minimum (522.06) re-verified 2026-09 directly against the version-specific archived release notes at docs.nvidia.com/cuda/archive/11.8.0/cuda-toolkit-release-notes — confirmed correct. The current live (non-archived) docs.nvidia.com/cuda/cuda-toolkit-release-notes page shows '520.06' in its cumulative Table 3 for this same row, which is an apparent error/staleness on NVIDIA's live page, not a second valid figure; the archived per-version page is authoritative for a shipped release. Last CUDA 11.x release; still the most common pin for older Pascal-era GPUs. computeCapabilities added 2026-08, cross-referenced against developer.nvidia.com/cuda-gpus and the CUDA C++ Programming Guide's compute-capabilities appendix; minor SKU-specific steppings (e.g. 8.7 for Jetson Orin, 12.1 for specific Blackwell SKUs) are folded into the primary value shown, matching this site's existing cuDNN '9.x'-style bucketing rather than listing every sub-variant.