The test runs a WebGPU compute shader for 30 seconds. Each element gets 256 dependent fused multiply-adds, and the result is reported as GFLOP/s.
- The headline is the median of one-second windows, with the spread between windows beside it, so a short dip does not move it.
- A second figure copies a 64 MiB buffer 20 times and reports the rate.
- It is a standalone test. It is not part of the Standard or Deep run and does not change the score.
The short answer
The WebGPU compute test asks your graphics chip to do arithmetic, not to draw. It measures how many floating-point operations per second the GPU completes through the browser's WebGPU interface, and how fast it copies a buffer in its own memory. Those are what a web app using WebGPU for compute can get on this machine.
How the figure is made
The shader loads a number, runs a chain of multiply-adds in which each step depends on the last, and stores the result, so the compiler cannot skip or merge the work. Each dispatch is submitted on its own and timed until the browser reports the GPU has finished it. A few warm-up dispatches are timed and dropped first.
The dispatches are grouped into one-second windows by their own summed times. The headline GFLOP/s is the median window, and the spread between the fastest and slowest windows is shown beside it. A power limit that moves during the run shows in the spread rather than in the headline.
What it does not measure
It is not a gaming figure. Games draw triangles and textures through the graphics pipeline, which the graphics stage of the main run exercises. The time per dispatch includes the browser's own submission and completion overhead, which is a real share on a fast GPU, and the page says so rather than subtracting a guess. The figure is not the GPU's rated peak, which counts every unit busy at full clock with no overhead.
WebGPU must be available in the browser. Where it is not, the test says so and does not run.
Terms used here
Questions
- Why is my GFLOP/s much lower than my GPU's specification?
- A specification sheet's figure is a theoretical peak. This test measures one dependent workload through a browser, including the browser's overhead, so it is expected to be lower. It is most useful compared with another run of the same test.
- Does the WebGPU test change my SystemCheck score?
- No. It runs on its own page, is not a stage of the Standard or Deep run, and contributes nothing to the score.