Why did one way of splitting Meta's 816B model run 59% faster on the same two Macs?

Why would two identical machines, running the same model, give you two different speeds? That is what Apple's machine learning team published on 15 October. They ran Muse Spark Open, the open-weights