What is a good setup for building medium size deep learning using Tensorflow with the Metal plugin

Question

Level 1

9 points

What is a good setup for building medium size deep learning using Tensorflow with the Metal plugin

I’m in the enviable position of having a nice wad of cash to spend on a new Mac setup. Since Tensorflow and PyTorch are now both supported by Metal, at least on my 16gb MacBook Air M1 I’m more than tempted by either a basic Mac Studio or a top end 16” MacBook with enough to add a Mac Studio monitor, mouse and keyboard on top. Both will take in excess of 48Gb and a 1Tb SSD. So here’s the rub which model and Ram combination is best? I could max out at 96gb or even 128gb but I suspect that would be wasteful. Would 48Gb on a MacBook 16” cut the mustard?

Posted on May 3, 2024 11:23 AM

Reply

Answer 1

May 3, 2024 1:30 PM in response to Nick Walton

Apple doesn’t have much guidance, but has posted a few details here:

https://machinelearning.apple.com/updates/ml-compute-training-on-mac

There are a few other articles:

https://medium.com/@kunsitu/tensorflow-performance-on-the-new-m1-max-macbook-8dee45649152

The closest I’ve found is a video:

https://www.youtube.com/watch?v=cpYqED1q6ro

I’d try asking here:

https://discuss.tensorflow.org/

A fair amount of what’s posted pre-dates the Apple Tensorflow Metal integration, too.

Some of this degenerates into “rent some big box in some cloud and run the training there”, when the requirements exceed local resources, too.

Reply