Mac M2 Max Tensor Matrix Algebra on GPU

[LopezRicardo1]

Is there a way to perform matrix algebra operations using Mac M2 Max Gpu system using Torch R?

[dfalbel]

That should just work if you move tensors to the MPS device.

Eg:

x <- torch_randn(10, 10)$to(device="mps")
y <- torch_randn(10, 10)$to(device="mps")

torch_mm(x, y)

If both tensors are on the MPS device, then the operation will happen on the GPU if there's a kernel for it.

[LopezRicardo1]

I see, thank you!

Though, using this setting the operation will be done in the GPU, right? How can I check if there's a kernel for GPU?

I'm still in a learning curve on this!

[dfalbel]

The easiest way is to run and see if it works :) At this point, most operations already support running on mac GPU's. If you have large enough tensors you will quicky see your GPU usage going up in the system monitor app:

library(torch)
x <- torch_randn(10000, 10000)$to(device="mps")

for (i in 1:1000) {
  y <- torch_mm(x, x)  
}
y[10]

[LopezRicardo1]

Yes, GPU 90% on system monitor!

Thanks!