Looking at chess results and how bad they are, you should take the MacBook Pro 16-inch M1 MAX with 32 GPU cores and 64 GB RAM and 8 TB SSD or buy the stronger M2 MAX / M2 Ultra MacBook.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Would the cheapest M1/M2 MBP be OK to run Stable Diffusion on MacOS...
- Thread starter GMGN
- Start date
- Sort by reaction score
Does Stable Diffusion use CoreML or Pytorch on Apple hardware?
If it does not use CoreML, it is normal for Stable Diffusion to be slow on Apple hardware because Pytorch has an experimental Metal backend. If Stable Diffusion is ported to Tensorflow, performance should be better because the Metal backend is more mature.
If it does not use CoreML, it is normal for Stable Diffusion to be slow on Apple hardware because Pytorch has an experimental Metal backend. If Stable Diffusion is ported to Tensorflow, performance should be better because the Metal backend is more mature.
What's the wattage pulled from wall under load?
9.23 seconds on laptop 3060 configured for 70W. 84W total system power pulled from wall. Maybe I'll get around to upgrading from Big Sur to Monterey to try on M1 MBA.
They all were warned well over a year ago to switch to Metal AND THAT YOU WILL BE VERY SORRY IF YOU DO NOT SWITCH TO METAL.
Well…feel sorry developers. You didn’t listen when it came to CodeWarrior vs XCode and you didn’t listen here, either. History tends to rhyme.
Do you think Pytorch and Tensorflow developers are like iOS developers who worry if their app doesn't run well on Apple hardware? Think again. Meta (Pytorch) and Google (Tensorflow) only care about running on large Nvidia GPUs, so Pytorch and Tensorflow have a very mature CUDA backend.
Only Apple benefits from Pytorch running fast on Apple hardware, so Apple could have given Pytorch a Metal backend earlier, as they have done with Tensorflow. If Apple cared about Pytorch, it would have joined the Pytorch foundation.
PyTorch Foundation
A lot of developers prefer the Mac experience and better support for these frameworks is a popular request. That’s also why PyTorch devs are pursuing Apple Silicon accelerated backend (if I understand correctly, with Apples assistance). It will still take a while for these implementations to mature and of course they won’t replace high-end Nvidia GPUs any time soon, but it’s something that’s being actively worked on.
And btw, I think the discussion here misses the point about viability of ML workloads on Apple Silicon. A MacBook doesn’t need to outperform an RTX 3080. The implementation just needs to be fast enough to support prototyping or ML assisted work. Serious inference will be done in the cloud anyway.
Why should Meta employees care more than Apple employees about whether Pytorch runs fast on Apple hardware? Apple hasn't even joined the Pytorch foundation. Besides, Pytorch and Tensorflow are open source projects, and anyone can improve them.
Tensorflow and Pytorch are developed primarily for data center GPUs, and any other hardware has lower priority. The Tensorflow developers were smart to decouple the kernel code from the backend, so they let third parties worry about other backends while they take care of their main focus, the backend for TPUs and Nvidia GPUs.
PluggableDevice: Device Plugins for TensorFlow
In this post, we introduce the PluggableDevice architecture which offers a plugin mechanism for registering devices with TensorFlow without the need
