WebGPU Benchmark

Test your GPU's WebGPU performance with two benchmarks: a matrix-multiply compute test and a triangle-rendering throughput test. Results include a combined score you can compare against typical hardware tiers.

How It Works

The benchmark runs two independent tests using the WebGPU API, then combines the results into a single score.

Compute Test

Two 512x512 matrices of 32-bit floats are allocated on the GPU. A WGSL compute shader multiplies them together using 8x8 workgroups (64 threads per workgroup, 4 096 workgroups total). After 2 warmup iterations, 10 timed iterations run back-to-back. The metric is effective GFLOPS:

GFLOPS = (10 × 2 × 512³ FLOPs) / elapsed_seconds / 1 000 000 000

This is a memory-bandwidth-bound workload using a naive (non-tiled) implementation, which makes it a practical measure of GPU memory subsystem throughput rather than peak shader throughput.

Render Test

10 000 random colored triangles are generated on the CPU once, then uploaded to a GPU vertex buffer. A simple WGSL vertex+fragment shader draws all triangles to a 640x480 canvas. The GPU submits as many frames as possible in 3 seconds. The metric is millions of triangles per second (M tri/s = FPS x 10 000 / 1 000 000).

Combined Score

score = Math.round(GFLOPS × 100 + (tris_per_sec / 10 000))

The compute test carries more weight (typically 70-80% of the score) because it stresses compute and memory bandwidth more uniformly across GPU architectures.