MP 7,1 Mac Pro 7,1 vs Mac Studio for ML / Data Science workloads

[mcs618]

Does anyone have any experience using a 2019 Mac Pro for machine learning / data science / Python workloads?

I know that I should just build a CUDA workstation but I’m stubborn and have many other reasons for using a Mac so I’d like to stick to macOS.

I know the M1 Ultra is fairly decent in this regard, and it’s fairly cost effective. But does the Mac Pro perform well enough to compete?

I want a Mac Pro because it’s upgradable, has a certain nostalgia to it (I know it’s only 3 years old but it feels nostalgic compared to current M1 powered Macs) and the ability to have huge amounts of RAM is important to me. And I could throw an Nvidia GPU in there and boot windows if I really had to. But I don’t want to compromise performance-wise compared to the M1 Ultra.

Does anyone have any thoughts? I wish I could test some scripts to compare the two machines.

 

[TrevorR90]

I'm a professional data scientist, I use my Mac Pro for all my work. CUDA isn't supported on the Mac side, so automatically, the Mac studio can be excluded. I have a 3090 ti on the bootcamp side that does tensor flow work.

The Mac Pro is perfect for me because I like to use windows for games/data science work etc. Then at the end of the day, I switch to the Mac side for my everyday stuff because like you said, the macOS is just superior to me.

 

[mattspace]

I'm a professional data scientist, I use my Mac Pro for all my work. CUDA isn't supported on the Mac side, so automatically, the Mac studio can be excluded. I have a 3090 ti on the bootcamp side that does tensor flow work.

The Mac Pro is perfect for me because I like to use windows for games/data science work etc. Then at the end of the day, I switch to the Mac side for my everyday stuff because like you said, the macOS is just superior to me.

I'm still fuzzy on the status of Nvidia cards in 2019 machines. When you're booted into macOS, does the presence of the Nvidia card interfere with anything?

 

[mcs618]

I'm still fuzzy on the status of Nvidia cards in 2019 machines. When you're booted into macOS, does the presence of the Nvidia card interfere with anything?

I’ve heard that it does. I think there’s an issue with kernel panics when the display goes to sleep? Something along those lines. I think there might be some sort of hacky fix for it though.

 

[mcs618]

I'm a professional data scientist, I use my Mac Pro for all my work. CUDA isn't supported on the Mac side, so automatically, the Mac studio can be excluded. I have a 3090 ti on the bootcamp side that does tensor flow work.

The Mac Pro is perfect for me because I like to use windows for games/data science work etc. Then at the end of the day, I switch to the Mac side for my everyday stuff because like you said, the macOS is just superior to me.

True. I know with the latest m1 optimized packages though the Mac Studio isn’t half bad as far as ML performance goes. Not at the level of CUDA but still decent.

Have you ever tried any projects on the macOS hardware in your Mac Pro? I’m sure it’s not close to a 3090 but I’m still curious how it performs for stuff that isn’t hardcore deep learning. Wondering what the specs of your machine are as well.

 

[mattspace]

I’ve heard that it does. I think there’s an issue with kernel panics when the display goes to sleep? Something along those lines. I think there might be some sort of hacky fix for it though.

Yeah, I'm very curious as to what @TrevorR90 is doing to keep it installed.

 

[DrEGPU]

I believe you can compile PyTorch for MacOS with Metal support, so theoretically your AMD GPU’s should work. I haven’t tried it out yet, but I might do it just for fun. I don’t think Macs are ready for serious data science work that “requires” GPU’s. The software stack is just not as mature as it is for Linux and Windows.

This assumes you use PyTorch, of course!

 

[mcs618]

I believe TensorFlow is also optimized for metal. This page shows a short explanation and a couple tests using an M1 MBP and the 2019 Mac Pro

Accelerating TensorFlow Performance on Mac

Accelerating TensorFlow 2 performance on Mac

blog.tensorflow.org