[jjcon]

That is on a extremely small network (from a 1998 paper) running on an extremely small dataset of small images (28px by 28px grayscale images) and they are comparing the ms per step which is going to vary dramatically. Not only that they do not bother to run the m1max and simply say it, "should run twice as fast as the M1 Pro"

In other words that benchmark is completely useless. They should run a standard network from the past decade at least (say VGG16) on useful image sizes and should give the 10 epoch training time if they want anything stable that may approximate hobbyist workloads.