PyTorch MPS is buggy. Even with the stable build. Something with cuda is far better imo. Personally I use a mbp 14’ with the M1 Pro base model for literally everything and then I have a desktop (had one cuz I play games, just upgraded the gpu to a cheap 3090 I found online, works like a charm for 99% of work loads when it comes to training something.

For the 1% when I do not have enough compute I use my universities cluster or compute Canada for distributed training.

Read this:

https://www.reddit.com/r/mac/comments/10gpu46/hardware_for_scientific_computing/j54jpj3/?utm_source=share&utm_medium=ios_app&utm_name=iossmf&context=3

My recommendation: (1) Abandon macOS and get a laptop with an Nvidia GPU. Or (2) If you don’t like working on Linux/Windows and prefer Macs, then get a cheap MBA and an laptop with an Nvidia GPU. Use the Mac for coding but run the codes over ssh on the Nvidia laptop. The combined price would not exceed that of a specced out MacBook Pro, while the perf benefit would be more than 10x. Or (3) If you want to both code and run your code on a Mac and also don’t want to carry two laptops, then get the highest specced MacBook Pro possible. Neural network training is computationally very expensive. Normally we run our neural networks in our lab servers that contain anywhere between 4 to 64 GPUs. Even the highest end M2 Maxes are nothing to an RTX 4090.

So today I found out that mps on PyTorch 13.1 (stable) has bugs causing a lack of learning. FashionMNIST accuracies bounced around in <10%. Switched to cpu and worked fine (>80%)

I’m just confused as to who the MacBook is for

none. run ur jobs on a pc, connect to it using vnc/remotedeskto a cheap laptop or if you are savvy, ssh.

much better setup and bang for buck imho.