File size: 2,893 Bytes
b94a1ed 7503a74 b94a1ed 7503a74 b94a1ed | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 | ---
pretty_name: "NVIDIA GPU CUDA compute capability by model"
license: other
license_name: source-terms-cited
license_link: https://referencesource.org/nvidia-cuda-compute-capability-by-gpu/
language:
- en
tags:
- reference-data
- provenance
- referencesource-org
configs:
- config_name: default
data_files:
- split: train
path: data/records.jsonl
---
# NVIDIA GPU CUDA compute capability by model
**Canonical, always-current version: https://referencesource.org/nvidia-cuda-compute-capability-by-gpu/**
Machine-readable: https://referencesource.org/nvidia-cuda-compute-capability-by-gpu/data.json — this mirror is a point-in-time copy.
- **Last verified:** 2026-10-01
- **Stale after:** 2027-01-29 (past this date, prefer the canonical copy —
it re-verifies on a cadence this snapshot does not)
- **Records:** 442
Which CUDA compute capability version (e.g. '8.9', '12.0') each NVIDIA GPU model supports, from NVIDIA's own compute-capability table. Compute capability determines which CUDA toolkit version a card needs, which `-arch=sm_XX` / `-gencode` build flags to pass (the target is the capability written without its dot, so 8.9 is sm_89 and 12.0 is sm_120), and whether a PyTorch/TensorFlow wheel built for one architecture will run on a given card. Answers 'what compute capability is my GPU', 'does CUDA 11.8 support the RTX 4090', 'what sm_XX target do I use for an A100'. NVIDIA's own page is the single source of truth and is already well-ranked for individual current-generation cards; the value here is completeness (legacy cards back to the earliest Fermi/Kepler-era GPUs, and every SKU in a generation, not just the flagship) in one machine-readable table, and being current the moment a new architecture (e.g. Blackwell, 12.x) ships, before third-party blogs catch up.
## Provenance — what every record carries
Every record carries `source` (the URL the value was read from) and
`source_quote` (a verbatim quote from that page stating it). A value you
cannot check against its page is indistinguishable from an invented one;
these can all be checked. Each record is also individually addressable on
the canonical site at its `url`.
## Licence position
Source is NVIDIA's own developer documentation (developer.nvidia.com). We take the fact (model to compute-capability mapping) plus attribution linking back; we do not reproduce NVIDIA's full page design or unrelated content.
Facts are not copyrightable; what this dataset takes from each source is a
fact plus a short attributed quote linking back to the page that states it.
## Sources
- https://developer.nvidia.com/cuda/gpus/legacy
- https://developer.nvidia.com/cuda/gpus
## Fields
See any record in `data/records.jsonl` — field names are self-describing,
and the canonical page for each record renders them with labels:
https://referencesource.org/nvidia-cuda-compute-capability-by-gpu/
|