WebGPU Compute Benchmark
Three fixed workloads measure browser GPU compute throughput. Same work, five trials and output checks on every run.
Run the test
Results · suite v1
| Workload | Output check | Median ms / dispatch |
|---|---|---|
| Floating-point vectors | — | — |
| Integer hashing | — | — |
| Matrix multiplication | — | — |
Score = 1000 ÷ geometric mean of the three median milliseconds. Compare only suite v1 scores. Timings include queue submission and completion, but exclude setup and readback; they are not hardware-only GPU timestamps.
What is measured?
262,144 values, 48 arithmetic rounds each.
262,144 integers, 64 mixing rounds each.
Two fixed 256 × 256 matrices multiplied.
Each task is warmed up, then run in five batches of four dispatches. The median batch time is divided by four. Four sample results are checked. Device choice, drivers, browser scheduling and power settings can change the score; WebGPU may use a software adapter. No data is uploaded.
Compare with the JavaScript benchmark · WebAssembly benchmark
About GPU computing
Graphics processors were designed to perform many similar operations at once. The same parallel approach can also accelerate work that does not draw an image, from numerical simulations to image processing. This is often called general-purpose GPU computing.
WebGPU gives compatible browsers access to modern graphics and compute capabilities through a standard web interface. Its compute workloads split a large problem into many smaller pieces that can run in parallel. Whether that helps depends on the problem, the device and the cost of moving data.
Benchmarks are controlled experiments, not a complete description of a computer. A short calculation can be influenced by power saving, background activity, cooling and browser scheduling. Comparing repeated runs with the same workload and browser conditions is more meaningful than treating a single number as a universal speed rating.