WebGPU and AI compatibility checker inspects the browser capabilities used by this site’s on-device AI tools. Run it explicitly to check WebAssembly, SIMD, Workers, shared-memory threads and an available WebGPU adapter.
The result maps basic prerequisites and possible CPU fallbacks to the existing AI catalogue. A short typed-array CPU task is measured in a worker. No model is downloaded, and neither this task nor adapter limits proves real inference speed or available VRAM.
How to use
- Open the page over HTTPS in the browser you want to check.
- Run the device check and inspect capabilities and reported limits.
- Follow the relevant tool link or save the diagnostic JSON for troubleshooting.
Frequently asked questions
Does a passing result guarantee an AI model works?
No. Model initialization, memory pressure, file size, drivers and language-pair availability still need a real run. The result only checks basic prerequisites.
Are WebGPU limits my VRAM size?
No. They describe allowed API allocations or workgroup sizes, not free graphics memory. The browser may also hide adapter information.
Do missing WASM threads block all tools?
No. This site has single-threaded WASM paths. Shared-memory threads generally also require cross-origin isolation.
Is the CPU task a GPU benchmark?
No. It times a fixed local typed-array calculation outside the interface. There is no inference score or automatic speed guarantee.
Availability is narrower than performance
The checker reuses the same bounded adapter and device probes as the model-download guard. Unavailable or redacted information is shown rather than inferred from a browser name.
The three current SIMD-only ONNX distributions need SIMD. GPU-assisted tools may use a slower CPU path, while the built-in translator also needs its browser API and a supported language pair.
Useful diagnostics without model downloads
The JSON records detected capabilities, selected adapter limits and catalogue decisions. No diagnostic upload is performed by this tool.
Compare results after browser or driver changes, but do not compare the small CPU timing as if it measured model inference. Actual input and device memory conditions still matter.
