A vendor-agnostic Vulkan backend cuts edge inference latency from 30 ms to 3 ms

ppchaos · reddit · 2026-07-29

A production edge setup for a video editing tool shows how to run ML inference across heterogeneous hardware without relying on CUDA.

The team uses ncnn’s Vulkan backend so the same stack works on NVIDIA, AMD, Intel integrated graphics, and Apple Silicon. On an RTX 4070 in fp16, ArcFace R50 drops from 30 ms on ONNX CPU to 3 ms on ncnn Vulkan, while SCRFD falls from 25 ms to 2.5 ms. Model storage also shrinks, with ArcFace going from 174 MB ONNX fp32 to 87 MB in ncnn fp16 weight storage. The main argument is not just speed: Vulkan exists on the machines they ship to, so users do not need vendor-specific runtimes or installs.

Original post →

More from Infra

Infra channel →