How it works
The test
- Your graphics card multiplies large matrices in WebGPU, the same kind of work a language model does.
- It warms up first, then runs in 32-bit floats (FP32) and, when the browser allows it, 16-bit (FP16).
- The whole run takes about 20 seconds. Nothing is downloaded or installed.
The score
0.6 × FP32 + 0.4 × FP16Score, in GFLOPS
When the browser can't run FP16, the score is FP32 GFLOPS alone.
Browser scores run lower than native tools. Compare cards here, not with other benchmarks.
Estimates
When you pick a card instead of testing, or the browser draws with a software renderer, we estimate from the card's rated FP32 speed. A browser reaches about a third of it. Estimates are marked as such and never go on the leaderboard.
The leaderboard
Every finished test adds one row. Scores are not verified, so runs far above what a card can do, or far outside its usual range, are held back from the board. Treat it as a rough guide. What a row contains
The name
Kasoku is Japanese for acceleration, which is what a graphics card does for AI.
Tiers
A card gets the lower of its score tier and its memory tier. A fast chip cannot run a model that does not fit in memory.
PotatoAny card
Tiny models and simple tools. Still more than nothing.
Starter250+ and 3 GB+
Small chat models, speech to text and background removal.
Solid900+ and 6 GB+
Everyday 3-4B chat models and quick image tools.
Strong2,400+ and 8 GB+
7-8B chat models at reading speed and local image generation.
Beast6,000+ and 16 GB+
Larger models, SDXL-class images and long context.