WebGPU (BCACHE shader) • WASM SIMD (6x8 FMA, futex thread pool) • tfjs WebGPU • tfjs WASM (XNNPACK) • f32
Results (GFLOPS — higher is better)
| Size |
numpyjs WebGPU |
numpyjs WASM |
tfjs WebGPU |
tfjs WASM |
Best |
WebGPU = GPU-resident f32 (no CPU readback). WASM = SIMD f32 (futex thread pool, or single-threaded fallback). tfjs WebGPU = kernel + await data(). tfjs WASM = XNNPACK multithreaded.