I think the most interesting benchmark of the M2’s potential advantage conventional Intel/Nvidia machines would be a memory-heavy ML training or inference work

I think the most interesting benchmark of the M2’s potential advantage conventional Intel/Nvidia machines would be a memory-heavy ML training or inference workload. Apple made some noises about having 190GB of unified memory as an advantage for that scenario, since the system wouldn’t need to wait for copying from COU RAM to GPU RAM but I’d like to see that put to the test. I’m also disappointed that Apple didn’t go for a 2-socket Mac Pro so they could offer a compute advantage over Mac Studio in addition to the PCI Express slots. Other than the IO, I can’t think of a reason to pick the Mac Pro.