Backend Comparison

Total Step Latency

Throughput

Cache Strategy

Total Step Latency

Throughput

What You Are Seeing

These charts measure how fast the Flappy World model generates tokens directly in your browser using ONNX Runtime.

Each point runs a short decode burst of 1, 4, 8, or 16 tokens. The first section compares WebGPU vs WASM, while the second compares KV-cache reuse against recomputing the context each step.