NVIDIA GPU CUDA compute capability by model
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.
The data
Showing the first 400 of 441 records. Every record has its own page, with its source and the quoted line that states it; these list the rest.
- Records 1–400GeForce 210* … Quadro P2000
- Records 401–441Quadro P2200 … Tesla S1070
Where this came from
Every record above links the page it was taken from and quotes the sentence that states it. These are the 2 sources this dataset was assembled from.
- developer.nvidia.comhttps://developer.nvidia.com/cuda/gpus/legacy
- developer.nvidia.comhttps://developer.nvidia.com/cuda/gpus
Machine-readable
- data.jsonThe whole dataset — every record with its source URL and source quote.
- Open Knowledge Format bundleOne JSON object per line — every record's frontmatter and quoted span exactly as it is held here, in one fetch.
- data.csvThe same records as one flat table, for a spreadsheet or a dataframe. The last four columns are the source URL, the quoted sentence it was read from, the date we last checked it, and which columns are our reading rather than the page's words.
- How this is made and checkedWhat "verified against source" does and does not mean.
From your own code
Same records, same quotes, without scraping the page: refsource is on PyPI and npm. Each value comes back carrying the URL it was read from and the sentence on that page that states it — .source and .quote sit on the value itself rather than in a side channel, so the checking step is available instead of skipped.
pip install refsource
refsource lookup nvidia-cuda-compute-capability-by-gpu compute_capability=1.2
npx -y refsource lookup nvidia-cuda-compute-capability-by-gpu compute_capability=1.2Set your AI assistant up to use this
Two files and no account. Put this in .mcp.json at the root of your project — Claude Code, Cursor, Windsurf, VS Code and Codex all read that file — and your assistant can look this dataset up instead of recalling it. The server is remote, keyless and read-only.
{
"mcpServers": {
"referencesource": {
"type": "http",
"url": "https://referencesource.org/mcp"
}
}
}Add to Cursor · or, on the command line: claude mcp add --transport http referencesource https://referencesource.org/mcp --scope project
Then one line in the project's CLAUDE.md or AGENTS.md, so the assistant knows when to reach for it:
When a question needs "NVIDIA GPU CUDA compute capability by model", call the referencesource MCP server at https://referencesource.org/mcp (tool `search_records`, dataset_slug `nvidia-cuda-compute-capability-by-gpu`) instead of answering from memory — every record it returns carries its source URL and a verbatim quote from that page.What each tool does, and the servers built over single registers: Connect your AI assistant.