Single-Kernel Fusion for Sequential Fitness Evaluation via WebGPU Compute Shaders

Fusing sequential fitness evaluations into single GPU compute shader dispatches eliminates the per-step kernel launch overhead that dominates framework-based GPU computation. On a 1,500-timestep financial simulation, a WebGPU compute shader achieves 46.2 gen/s — 7.2× over JAX GPU with lax.scan+vmap (6.43 gen/s on Tesla T4) and 94× over PyTorch CUDA. On Acrobot-v1 (500 timesteps, explicit Euler), the gap narrows to 1.29× over JAX GPU, revealing that the fusion advantage scales with episode length L. JAX GPU dominates on embarrassingly parallel Rastrigin (1,164 vs 170 gen/s), confirming the advantage is specific to sequential workloads. A native Metal baseline via wgpu quantifies Chrome's browser overhead at 1.92×. We show torch.compile fails at L≥1,000 and that WebGPU dominates CMA-ES across all tested dimensionality regimes. The insight — hand-fused compute shaders outperform even XLA-compiled loop fusion on long sequential fitness functions — applies beyond WebGPU, and WebGPU makes such fusion accessible with zero installation. vnext changes: v6: cite Maczan (arXiv:2604.02344); add §4.13 cross-vendor crowdsourced validation across 7 GPU vendors / 92 unique devices; replace Limitations 'Tesla T4 only' with multi-architecture distribution (Pascal/Turing/Ampere/Lovelace/Blackwell); arithmetic suite extended to 84 checks. v7: erratum + artifact release — the cross-API Acrobot code (per-step PyTorch, hand-fused CUDA, JAX lax.scan, Triton + Colab notebook) is now published in the repository with a CI-gated cross-implementation equivalence check (NumPy/PyTorch/JAX identical in f64); integrator label corrected (explicit Euler dt=0.05 across all implementations, not RK4); §4.13 crowdsourced snapshot recovered into the repo (680 runs, exact match); re-measured on Colab T4 with verified-equivalent kernels (torch 2.11.0: fused CUDA 176x, JAX 33.5x over per-step PyTorch, which improved ~3x since the original measurement — ratios compress as frameworks improve, qualitative result unchanged).