Same data.
Different runtime.
Real CUDA-style kernels, compiled into browser compute shaders. Three.js renders the same GPU buffers.
.cuGPU0 copies**No compute-to-render data copy or CPU readback. Initial state and small scalar parameters are uploaded.
This supports a defined CUDA C kernel subset. It does not execute CUDA binaries, cuBLAS, inline PTX, arbitrary pointers or warp-specific instructions.
Kernel workbench
Compile a bundled kernel or open your own source. GPU validation does not prove your algorithm is correct.
The block shape must match shared-tile assumptions in your source. Use the JavaScript API to bind and launch custom kernels.
Test the actual GPU.
Independent references, empty inputs, odd sizes, tail guards, shared-memory barriers and integer atomics. Includes a rendered-pixel interop test.
No simulated or prefilled GPU results.
These tests cover the bundled kernels, not arbitrary edited source. Rendering pauses during the test run.
Run the official NVIDIA simpleGL kernel showcase ↗
Open measured RTX 5080 CUDA ↔ WebGPU comparison ↗
Find the fast path.
Sweep 64 / 128 / 256-lane scalar and float4 kernels. Compare ordinary, shared-tiled and register-tiled matrix multiplication.
Compilation and initial upload are excluded.
Every variant is checked before it is timed.
Reported winners apply to this adapter and workload size. Faster than another WebGPU kernel is not proof of CUDA parity.