Headroom: browser tool measures your GPU's true bandwidth ceiling for local AI

Headroom is a WebGPU-based browser tool that measures a device's real memory-bandwidth ceiling for local AI inference — the figure that ultimately caps tokens-per-second throughput. Rather than downloading actual models, it uses synthetic weights and fetches only about 0.4 MB of public tensor metadata, requiring no account, ads, or telemetry; all results stay local until a user opts to share them.

The tool also flags a known WebGPU subgroup driver bug that causes some in-browser LLMs to silently produce garbage output, a subtle failure mode engineers might otherwise miss. Measurements are validated against an FP64 reference computed on the CPU before timing runs, and projected tok/s ceilings are shown for model classes like Gemma variants and 8B q4 configurations.

A key finding highlighted by the tool: real inference engines typically achieve only 30-50% of the theoretical bandwidth ceiling, with the gap attributed to engine overhead. Validated on M1 and RTX 5070 hardware, Headroom requires WebGPU support (desktop Chrome/Edge or recent Android Chrome) and offers developers a concrete, zero-install benchmark for understanding hardware limits when building or tuning local LLM inference.