Comment by fancyfredbot

The write up on the GitHub repo is much more informative than the blog.

When running int8 matmul using onnx performance is ~0.6TF.

https://github.com/usefulsensors/qc_npu_benchmark