This category is for any question related to MPS support on Apple hardware (both M1 and x86 with AMD machines).
Intel GPUs were not supported by Apple’s TensorFlow backend, even though MPS and MPS Graph supports Intel GPUs with exceptionally high ALU utilization for matrix multiplications. I have a Mac mini with an Intel GPU, and other people stuck with older Macs may not have access to an Apple or AMD GPU. Intel Macs don’t have a 1.0 TFLOPS AMX accelerator on their CPU, so the GPU is the fastest part of the chip. Is there any chance the PyTorch backend will support them in the future?
The current state of the intel based mac is as follows:
- x86 macs with intel GPUs are not supported by MPS
at all (on the Apple side). So this won’t be supported I’m afraid.(EDIT: Sorry, this is not accurate, let’s continue the discussion about this in PyTorch support for Intel GPUs on Mac )
- x86 macs with AMD GPUs are supported by MPS and that should work on master now. We have some issues with fixing up final binaries for it. So if you need it now, you can compile from source. Otherwise, we expect to get the nightly binaries that support this later this week (you can track [DO NOT MERGE] Move x86 binaries builder to macos-12 to enable MPS build by albanD · Pull Request #77662 · pytorch/pytorch · GitHub ).
btw: this thread is only for the category specification discussion. Feel free to open a new topic if you want more details on this!
I have proof that says otherwise. Under the “Intel UHD 630” file, I have MPS running matrix multiplications with up to 95% ALU utilization.
Edit: I agree that further discussion should move into another category ( PyTorch support for Intel GPUs on Mac ), sorry for not reading your last edit before posting.