Does onnxruntime on macOS with apple silicon uses arm-specific instructions when running quantized models?

[Rexhaif]

Hi,

I'm doing some research with quantized models and i want to confirm one thing.

Does onnxruntime python package, installed on MacOS with Apple Silicon, actually uses arm-specific instructions(like faster int8 matmul) when running inference of quantized onnx models?

Specifically, i'm using onnxruntime through optimum library, and i am running BERT-base model, dynamically quantized for arm64 with this config(i.e u8/s8 quantization). My hardware is MBP 13 with Apple M1, macOS 13.2.1.