{ "dataset": "cuda", "source": "https://gpucompat.com/data/cuda.json", "about": "https://gpucompat.com/about-the-data/", "count": 21, "entries": [ { "id": "10-0", "version": "10.0", "releaseDate": "2018-09 (10.0.130 GA)", "status": "eol", "minDriverLinux": "410.48", "minDriverWindows": "411.31", "supportedArchitectures": [ "Kepler", "Maxwell", "Pascal", "Volta", "Turing" ], "computeCapabilities": [ "3.0", "3.2", "3.5", "3.7", "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5" ], "cudnnCompatible": [ "7.x" ], "wslDockerNotes": "WSL2: CUDA-on-WSL2 GPU support didn't exist yet when CUDA 10.0 shipped -- NVIDIA introduced it circa 2020, requiring a Windows driver from the R495 branch or later (docs.nvidia.com/cuda/wsl-user-guide). Using CUDA under WSL2 today means installing a current CUDA Toolkit release, not this one. Docker: independent of WSL2 -- a host with a driver >= the Linux minimum above and the NVIDIA Container Toolkit installed can still run this CUDA version inside a container; the CUDA Toolkit itself only needs to be present in the container image, not on the host.", "notes": "Added 2026-09 at the user's request to extend coverage back one step further than the existing 10.2 entry. Driver minimums (Linux 410.48 / Windows 411.31) confirmed via a driver-requirements table reproduced in a later CUDA Toolkit release notes PDF (NVIDIA's own 10.0 release notes archive was not directly fetched). Release date/month per contemporaneous coverage (Phoronix, 2018-09) of CUDA 10's launch alongside Turing support — exact day not verified. This was the first CUDA release to support Turing (compute capability 7.5), announced alongside the RTX 20 series. Unlike the 10.2/11.x entries on this site, which start their architecture list at Maxwell, this entry includes Kepler (3.0/3.2/3.5/3.7): Kepler support wasn't deprecated until 10.2 and wasn't dropped until 11.0, so 10.0 is a genuine, not inferred, match for the K80's targetId of 3.7 — see the note on nvidia-k80.yaml, which will need updating since it currently claims no CUDA entry on this site lists 3.7. Not independently cross-checked against NVIDIA's per-version architecture support matrix beyond the Turing-launch fact above." }, { "id": "10-1", "version": "10.1", "releaseDate": "2019-02 (10.1.105 GA)", "status": "eol", "minDriverLinux": "418.39", "minDriverWindows": "418.96", "supportedArchitectures": [ "Kepler", "Maxwell", "Pascal", "Volta", "Turing" ], "computeCapabilities": [ "3.0", "3.2", "3.5", "3.7", "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5" ], "cudnnCompatible": [ "7.x" ], "wslDockerNotes": "WSL2: CUDA-on-WSL2 GPU support didn't exist yet when CUDA 10.1 shipped -- NVIDIA introduced it circa 2020, requiring a Windows driver from the R495 branch or later (docs.nvidia.com/cuda/wsl-user-guide). Using CUDA under WSL2 today means installing a current CUDA Toolkit release, not this one. Docker: independent of WSL2 -- a host with a driver >= the Linux minimum above and the NVIDIA Container Toolkit installed can still run this CUDA version inside a container; the CUDA Toolkit itself only needs to be present in the container image, not on the host.", "notes": "Added 2026-09 alongside 10.0 at the user's request. Driver minimums (Linux 418.39 / Windows 418.96) are the 10.1.105 GA figures; NVIDIA's own 10.1 release notes archive shows later 10.1.x point releases (e.g. 10.1.168) raised the minimum slightly (Linux driver 418.67), but this entry tracks the GA baseline, matching this site's minor-version granularity elsewhere (see the 11.2 entry's equivalent caveat). Release month per Microsoft's C++ team blog post announcing CUDA 10.1 (Feb/Mar 2019); exact GA day not verified. Architecture/compute-capability list carried over unchanged from the 10.0 entry — CUDA 10.1 was a point release with no new architecture support — including Kepler (3.0/3.2/3.5/3.7), which this site's 10.2-and-later entries omit since Kepler was deprecated in 10.2 and dropped in 11.0." }, { "id": "10-2", "version": "10.2", "releaseDate": "2019 (exact month unverified — sources disagree on whether this refers to the original GA or a later 10.2.89 update)", "status": "eol", "minDriverLinux": "440.33", "minDriverWindows": "441.22", "supportedArchitectures": [ "Kepler", "Maxwell", "Pascal", "Volta", "Turing" ], "computeCapabilities": [ "3.0", "3.2", "3.5", "3.7", "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5" ], "cudnnCompatible": [ "7.x" ], "wslDockerNotes": "WSL2: CUDA-on-WSL2 GPU support didn't exist yet when CUDA 10.2 shipped -- NVIDIA introduced it circa 2020, requiring a Windows driver from the R495 branch or later (docs.nvidia.com/cuda/wsl-user-guide). Using CUDA under WSL2 today means installing a current CUDA Toolkit release, not this one. Docker: independent of WSL2 -- a host with a driver >= the Linux minimum above and the NVIDIA Container Toolkit installed can still run this CUDA version inside a container; the CUDA Toolkit itself only needs to be present in the container image, not on the host.", "notes": "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). Verified 2026-08 against NVIDIA's CUDA Toolkit release notes (Table 3) for driver minimums. Predates Ampere (added in CUDA 11.0). Added to fix a dead cross-link: PyTorch 1.13.0's page listed this version as a supported CUDA build before this entry existed. 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." }, { "id": "11-2", "version": "11.2", "releaseDate": "2020-12 (approximate — not independently confirmed from a single authoritative source)", "status": "eol", "minDriverLinux": "460.27.03", "minDriverWindows": "460.82", "supportedArchitectures": [ "Kepler", "Maxwell", "Pascal", "Volta", "Turing", "Ampere" ], "computeCapabilities": [ "3.5", "3.7", "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6" ], "cudnnCompatible": [ "8.x" ], "wslDockerNotes": "WSL2: this version's own driver minimum above (Linux 460.27.03 / Windows 460.82) is below the R495 driver branch NVIDIA's CUDA-on-WSL2 guide requires for GPU support under WSL2 (docs.nvidia.com/cuda/wsl-user-guide) -- meeting the minimum in this entry is not enough to get this CUDA version working inside WSL2; install a Windows driver from the R495 branch or later instead, and never 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.", "notes": "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). Verified 2026-08 against NVIDIA's CUDA Toolkit release notes (Table 3) for the 11.2.0 GA driver minimums; later 11.2.x update releases (11.2.1/11.2.2) raised the minimums slightly (Linux 460.32.03; Windows 461.09/461.33) but this entry tracks the GA baseline, matching this site's minor-version granularity elsewhere. Added to fix dead cross-links from the tensorflow collection (2.6.0 through 2.11.0 all pin CUDA 11.2). Architecture/compute-capability list is inferred from the 11.6 entry (same Maxwell-through-Ampere range; Ada Lovelace/Hopper weren't introduced until 11.8/12.0), not independently re-verified against NVIDIA's per-version architecture matrix for 11.2 specifically." }, { "id": "11-3", "version": "11.3", "releaseDate": "2021-04 (approximate — Phoronix/DEVCLASS coverage dates the 11.3.0 GA to mid-April 2021, not independently confirmed from a single authoritative NVIDIA source)", "status": "eol", "minDriverLinux": "465.19.01", "minDriverWindows": "465.89", "supportedArchitectures": [ "Kepler", "Maxwell", "Pascal", "Volta", "Turing", "Ampere" ], "computeCapabilities": [ "3.5", "3.7", "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6" ], "cudnnCompatible": [ "8.x" ], "wslDockerNotes": "WSL2: this version's own driver minimum above (Linux 465.19.01 / Windows 465.89) is below the R495 driver branch NVIDIA's CUDA-on-WSL2 guide requires for GPU support under WSL2 (docs.nvidia.com/cuda/wsl-user-guide) -- meeting the minimum in this entry is not enough to get this CUDA version working inside WSL2; install a Windows driver from the R495 branch or later instead, and never 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.", "notes": "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). Added 2026-09 to fix a dead cross-link: adding PyTorch 1.12.0 (which lists CUDA 11.3 as one of its official wheel builds) would otherwise link to a nonexistent page, the same gap this site's 10.2/11.2/11.6 entries were each added to close for other PyTorch/TensorFlow versions. Driver minimums (Linux 465.19.01 / Windows 465.89) confirmed via NVIDIA's CUDA Toolkit release notes (Table 3), consistent across the 11.3.0 GA and 11.3.1 update. Architecture/compute-capability list carried over unchanged from the 11.2/11.6 entries (same Maxwell-through-Ampere range) rather than independently re-verified against NVIDIA's per-version architecture matrix for 11.3 specifically." }, { "id": "11-6", "version": "11.6", "releaseDate": "2022-01 (approximate — not independently confirmed from a single authoritative source)", "status": "eol", "minDriverLinux": "510.39.01", "minDriverWindows": "511.23", "supportedArchitectures": [ "Kepler", "Maxwell", "Pascal", "Volta", "Turing", "Ampere" ], "computeCapabilities": [ "3.5", "3.7", "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6" ], "cudnnCompatible": [ "8.x" ], "wslDockerNotes": "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.", "notes": "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). Verified 2026-08 against NVIDIA's CUDA Toolkit release notes (Table 3) for driver minimums. Added to fix a dead cross-link: PyTorch 1.13.0's page listed this version as a supported CUDA build before this entry existed. 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." }, { "id": "11-7", "version": "11.7", "releaseDate": "2022-05 (approximate — not independently confirmed from a single authoritative source)", "status": "eol", "minDriverLinux": "515.43.04", "minDriverWindows": "516.01", "supportedArchitectures": [ "Kepler", "Maxwell", "Pascal", "Volta", "Turing", "Ampere" ], "computeCapabilities": [ "3.5", "3.7", "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6" ], "cudnnCompatible": [ "8.x" ], "wslDockerNotes": "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.", "notes": "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). Verified 2026-08 against NVIDIA's CUDA Toolkit release notes (Table 3) for driver minimums. Added to fix dead cross-links: PyTorch 1.13.0's and 2.0.0's pages both listed this version as a supported CUDA build before this entry existed. 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." }, { "id": "11-8", "version": "11.8", "releaseDate": "2022-10", "status": "supported", "minDriverLinux": "520.61.05", "minDriverWindows": "522.06", "supportedArchitectures": [ "Kepler", "Maxwell", "Pascal", "Volta", "Turing", "Ampere", "Ada Lovelace", "Hopper" ], "computeCapabilities": [ "3.5", "3.7", "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6", "8.9", "9.0" ], "cudnnCompatible": [ "8.x" ], "wslDockerNotes": "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.", "notes": "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." }, { "id": "12-0", "version": "12.0", "releaseDate": "2022-12", "status": "supported", "minDriverLinux": "525.60.13", "minDriverWindows": "527.41", "supportedArchitectures": [ "Maxwell", "Pascal", "Volta", "Turing", "Ampere", "Ada Lovelace", "Hopper" ], "computeCapabilities": [ "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6", "8.9", "9.0" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "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). Verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. First 12.x release; first CUDA major-version bump since 11.0. 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." }, { "id": "12-1", "version": "12.1", "releaseDate": "2023-02", "status": "supported", "minDriverLinux": "530.30.02", "minDriverWindows": "531.14", "supportedArchitectures": [ "Maxwell", "Pascal", "Volta", "Turing", "Ampere", "Ada Lovelace", "Hopper" ], "computeCapabilities": [ "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6", "8.9", "9.0" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "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). Verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. 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." }, { "id": "12-2", "version": "12.2", "releaseDate": "2023-06", "status": "supported", "minDriverLinux": "535.54.03", "minDriverWindows": "536.25", "supportedArchitectures": [ "Maxwell", "Pascal", "Volta", "Turing", "Ampere", "Ada Lovelace", "Hopper" ], "computeCapabilities": [ "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6", "8.9", "9.0" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "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). Verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. 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." }, { "id": "12-3", "version": "12.3", "releaseDate": "2023-10", "status": "supported", "minDriverLinux": "545.23.06", "minDriverWindows": "545.84", "supportedArchitectures": [ "Maxwell", "Pascal", "Volta", "Turing", "Ampere", "Ada Lovelace", "Hopper" ], "computeCapabilities": [ "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6", "8.9", "9.0" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "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). Verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. 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." }, { "id": "12-4", "version": "12.4", "releaseDate": "2024-03", "status": "supported", "minDriverLinux": "550.54.14", "minDriverWindows": "551.61", "supportedArchitectures": [ "Maxwell", "Pascal", "Volta", "Turing", "Ampere", "Ada Lovelace", "Hopper" ], "computeCapabilities": [ "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6", "8.9", "9.0" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "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). Verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. 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." }, { "id": "12-5", "version": "12.5", "releaseDate": "2024-05", "status": "supported", "minDriverLinux": "555.42.02", "minDriverWindows": "555.85", "supportedArchitectures": [ "Maxwell", "Pascal", "Volta", "Turing", "Ampere", "Ada Lovelace", "Hopper" ], "computeCapabilities": [ "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6", "8.9", "9.0" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "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). Verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. 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." }, { "id": "12-6", "version": "12.6", "releaseDate": "2024-08", "status": "supported", "minDriverLinux": "560.28.03", "minDriverWindows": "560.76", "supportedArchitectures": [ "Maxwell", "Pascal", "Volta", "Turing", "Ampere", "Ada Lovelace", "Hopper" ], "computeCapabilities": [ "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6", "8.9", "9.0" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "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). Verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. GA figures shown; Update 1-3 sub-releases bump the minimum driver slightly higher (e.g. Update 3 requires Linux >=560.35.05 / Windows >=561.17) — see official release notes for the exact Update-level table if pinning to a specific patch. 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." }, { "id": "12-8", "version": "12.8", "releaseDate": "2025-01", "status": "supported", "minDriverLinux": "570.26", "minDriverWindows": "570.65", "supportedArchitectures": [ "Maxwell", "Pascal", "Volta", "Turing", "Ampere", "Ada Lovelace", "Hopper", "Blackwell" ], "computeCapabilities": [ "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6", "8.9", "9.0", "10.0", "12.0" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "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). Driver minimums verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. First CUDA release with Blackwell architecture support, per the official cuDNN support matrix. Release date (January 2025, for the 12.8.0 GA point release) verified 2026-09 against developer.nvidia.com/cuda-toolkit-archive, NVIDIA's own toolkit archive listing. 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." }, { "id": "12-9", "version": "12.9", "releaseDate": "2025-05", "status": "supported", "minDriverLinux": "575.51.03", "minDriverWindows": "576.02", "supportedArchitectures": [ "Maxwell", "Pascal", "Volta", "Turing", "Ampere", "Ada Lovelace", "Hopper", "Blackwell" ], "computeCapabilities": [ "5.0", "5.2", "5.3", "6.0", "6.1", "6.2", "7.0", "7.2", "7.5", "8.0", "8.6", "8.9", "9.0", "10.0", "10.3", "12.0", "12.1" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "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). Driver minimums verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. Release date (May 2025, for the 12.9.0 GA point release) verified 2026-09 against developer.nvidia.com/cuda-toolkit-archive, NVIDIA's own toolkit archive listing. 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. Compute capabilities 10.3 (B300) and 12.1 (GB10 / DGX Spark) added 2026-09-27: the CUDA 12.9 release notes say it adds compiler target support for SM architecture 10.3 (sm_103, sm_103f, sm_103a) and 12.1 (sm_121); 13.x keeps all architectures from Turing through Grace Blackwell." }, { "id": "13-0", "version": "13.0", "releaseDate": "2025-08", "status": "supported", "minDriverLinux": "580.65.06", "minDriverWindows": "N/A — not bundled with the Toolkit starting with CUDA 13.0 itself (users install the driver separately)", "supportedArchitectures": [ "Turing", "Ampere", "Ada Lovelace", "Hopper", "Blackwell" ], "computeCapabilities": [ "7.5", "8.0", "8.6", "8.9", "9.0", "10.0", "10.3", "12.0", "12.1" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "Driver minimums verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. First 13.x release; Pascal and Volta support dropped starting this version, per the official CUDA Toolkit/Driver/Architecture matrix (both listed as 'last support: CUDA 12.x'). Release date (August 2025, for the 13.0.0 GA point release) verified 2026-09 against developer.nvidia.com/cuda-toolkit-archive. Windows driver-bundling change re-verified 2026-09 directly against the version-specific archived release notes (docs.nvidia.com/cuda/archive/13.0.0/cuda-toolkit-release-notes): the 'no longer bundled' change starts with 13.0 itself, not 13.1 as a prior pass on this file had it — corrected. 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. Compute capabilities 10.3 (B300) and 12.1 (GB10 / DGX Spark) added 2026-09-27: the CUDA 12.9 release notes say it adds compiler target support for SM architecture 10.3 (sm_103, sm_103f, sm_103a) and 12.1 (sm_121); 13.x keeps all architectures from Turing through Grace Blackwell." }, { "id": "13-1", "version": "13.1", "releaseDate": "2025-12", "status": "supported", "minDriverLinux": "590.44.01", "minDriverWindows": "N/A — Windows display driver no longer bundled with the CUDA Toolkit package starting this version (install separately)", "supportedArchitectures": [ "Turing", "Ampere", "Ada Lovelace", "Hopper", "Blackwell" ], "computeCapabilities": [ "7.5", "8.0", "8.6", "8.9", "9.0", "10.0", "10.3", "12.0", "12.1" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "Driver minimums verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. Release date (December 2025, for the 13.1.0 GA point release) verified 2026-09 against developer.nvidia.com/cuda-toolkit-archive. 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. Compute capabilities 10.3 (B300) and 12.1 (GB10 / DGX Spark) added 2026-09-27: the CUDA 12.9 release notes say it adds compiler target support for SM architecture 10.3 (sm_103, sm_103f, sm_103a) and 12.1 (sm_121); 13.x keeps all architectures from Turing through Grace Blackwell." }, { "id": "13-2", "version": "13.2", "releaseDate": "2026-03", "status": "supported", "minDriverLinux": "595.45.04", "minDriverWindows": "N/A — Windows display driver not bundled with the CUDA Toolkit package (install separately)", "supportedArchitectures": [ "Turing", "Ampere", "Ada Lovelace", "Hopper", "Blackwell" ], "computeCapabilities": [ "7.5", "8.0", "8.6", "8.9", "9.0", "10.0", "10.3", "12.0", "12.1" ], "cudnnCompatible": [ "9.x" ], "wslDockerNotes": "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.", "notes": "Driver minimums verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08. Release date (March 2026, for the 13.2.0 GA point release) verified 2026-09 against developer.nvidia.com/cuda-toolkit-archive. 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. Compute capabilities 10.3 (B300) and 12.1 (GB10 / DGX Spark) added 2026-09-27: the CUDA 12.9 release notes say it adds compiler target support for SM architecture 10.3 (sm_103, sm_103f, sm_103a) and 12.1 (sm_121); 13.x keeps all architectures from Turing through Grace Blackwell." }, { "id": "13-3", "version": "13.3", "releaseDate": "2026-05", "status": "supported", "minDriverLinux": "610.43.02", "minDriverWindows": "N/A — Windows display driver not bundled with the CUDA Toolkit package (install separately)", "supportedArchitectures": [ "Turing", "Ampere", "Ada Lovelace", "Hopper", "Blackwell" ], "computeCapabilities": [ "7.5", "8.0", "8.6", "8.9", "9.0", "10.0", "10.3", "12.0", "12.1" ], "cudnnCompatible": [ "9.24.0" ], "wslDockerNotes": "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.", "notes": "Driver minimums verified against NVIDIA CUDA Toolkit release notes (Table 3), 2026-08 — current latest CUDA release as of this data pull. cuDNN 9.24.0 is the recommended pairing per NVIDIA's own tuning-heuristics guidance. Release date (May 2026, for the 13.3.0 GA point release) verified 2026-09 against developer.nvidia.com/cuda-toolkit-archive. 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. Compute capabilities 10.3 (B300) and 12.1 (GB10 / DGX Spark) added 2026-09-27: the CUDA 12.9 release notes say it adds compiler target support for SM architecture 10.3 (sm_103, sm_103f, sm_103a) and 12.1 (sm_121); 13.x keeps all architectures from Turing through Grace Blackwell." } ] }