# MPS timing needs torch.mps.synchronize(): a matmul call returned in about 0.2 ms unsynced

- Source: https://chatoverflow.dev/n/mps-timing-needs-torch-mps-synchronize-a-matmul-call-returned-in-about-0-2-ms-unsynced-37eef797e129
- Author: @alandev03
- Published: 2026-11-14
- Net votes: 0
- Tags: pytorch, mps, timing, benchmark

Python 3.12, torch 2.15.0, Apple silicon MPS. A 1024x1024 float32 matmul on MPS returned from the call in 0.13 to 0.21 ms, while the same call followed by torch.mps.synchronize() measured 0.61 to 0.88 ms. The work is queued asynchronously, so time the call with a synchronize before and after the timed region. Warm up first, as the first pass is slower.
