M4 Pro achieves 8ms realtime inference for monocular depth model

fofrAI · x · 2026-08-18

A developer optimized a 448x448 monocular depth model to run in 8ms on an M4 Pro across 250 dispatches, making it fast enough for real-time use. Written directly in TypeGPU, the inference feeds the depth buffer straight into the lighting pass without leaving the GPU or extra synchronization steps. Inference, lighting, and drawing all go through the same command encoder.

Original post →

More from Infra

Infra channel →