Home▸Computer Science▸Inference Optimization & Compression: Quantization, Pruning & Speculative Decoding

🗜️ Inference Optimization & Compression: Quantization, Pruning & Speculative Decoding

Watch a transformer shrink and speed up: switch FP32/FP16/INT8 quantization, prune weak connections, and enable speculative decoding to see model size, latency and throughput change live.

Computer Science3DModerate60 FPS

Details

CategoryComputer Science

RenderingThree.js · WebGL

DifficultyModerate

Target FPS60

ControlsPrecision buttons (FP32/FP16/INT8) set bit-width, node size/jitter and per-op speed; pruning-sparsity slider (0-90%) thins weak connections; draft+verify toggle turns speculative decoding on/off; draft-length slider (gamma, 1-8) sets tokens proposed per verification round; Pause/Reset.

Three.jsquantizationpruningspeculative decodingLLM inferenceInstancedMesh

Related simulations

Hash Function Avalanche VisualizerInteractive 3D bit-grid where flipping a single input bit shows how a cryptographic hash's…

Merkle Tree Proof VisualizerInteractive 3D binary hash tree where selecting a leaf and generating its proof path shows…

CPU Scheduling Policy SimulatorInteractive 3D process queue where switching between FCFS, SJF, priority and round-robin… ⊞ All 83 in Computer Science →

⚙ Under the hood

Switch a transformer between FP32/FP16/INT8 quantization, prune weak connections and toggle speculative decoding to watch model size, latency and throughput change live.

Three.jsquantizationpruningspeculative decodingLLM inferenceInstancedMesh

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install