The four public result states
| State | What it means | What it never means |
|---|---|---|
| Supported | Every required checked dimension has a positive scoped rule. | A universal stability or performance guarantee. |
| Unsupported | A required dimension has an authoritative negative rule. | That every other configuration is unsupported. |
| Conditional | Some rules pass, but prerequisites, provider rules, application versions, or lifecycle caveats still matter. | That the reader can skip local validation. |
| Unknown | Coverage is incomplete, stale, unavailable, or conflicting. | That the driver is bad or good. |
Inputs a real lookup must preserve
A useful result needs the exact GPU model, OS and build, CPU architecture, desktop/notebook/server or provider context, installed or proposed driver, and workload. Advanced cases may need CUDA runtime/toolkit, library, WSL, container, VM, Linux distribution, kernel, module flavor, Secure Boot, or application version.
The existing CUDA compatibility guide remains the canonical explanation for driver, Toolkit, runtime, PTX, and framework-layer confusion.
[33][34][60][61]No automatic safe-upgrade verdict
The hub can identify eligible or blocked facts. It cannot promise that an unknown system should upgrade, roll back, disable security controls, or run cleanup utilities. Installation recovery belongs to source-reviewed procedures and exact package scope.
[70][71][72]