1

I recently saw on MacRumors that Apple has began selling on their Clearance & Refurbished marketplace the 2021 MacBook Pro models. These machines feature the M1 Chips with X CPUs and Y GPUs.

I am a Data Scientist and a Mac lover. I'm also entering the Deep Learning field and I was wondering whether such machines are good for training neural nets or XGBoost models.

I have read on various forums that, despite the GPUs and the Neural Engine hype, Apple is only gradually releasing updates to allow the use of, eg. TensorFlow, on its M1 chip machines.

Given such premises, would you recommend replacing my present 2018 MacBook with a refurbished, 2021 M1 model?

Improve this question

1 Answer 1

Reset to default
2

I can’t benchmark your workload in detail, but TensorFlow should have good official coverage from Apple running TensorFlow on Apple Silicon / Metal.

Also I would trade up in a heartbeat from the machines you list. Here is a MacBook Air handing it to two Intel “workstations” in a benchmark. (3x to 4x speedup)

enter image description here

I love Apple refurbished - it saves me the cost of AppleCare typically and the quality has been stellar in my years of seeing them ordered.

In your case, the 2018 model (as will most Intel based Macs except the iMac Pro and Mac Pro) will likely decline in value pretty quickly now and the M1 will likely hold resale value for much longer. You might even trade it in with Apple for an easy $ if you don’t want to reuse or resell it yourself.

The code for triald and CoreML is still being written and will be much slower on the Intel CPU vs Apple Silicon

You will also get battery life benefits and other features as part of the upgrade. Unless your libraries and models are very Intel specific, you’ll want to be where Apple is investing all of its engineering which is the iOS / iPadOS and macOS SDK for Apple Silicon.

Due to drivers and GPU you’ll want an iMac Pro or Mac Pro and likely do heavy lifting in Windows for CUDA work. Converting to CoreML should be fine as is using either MacBook Pro or any M1 (even the air) to refine the models or if you’re not doing this professionally and can not fund a Xeon class workstation ($4k and up US).

You can have a super snappy Mac with all your features you love for a discount. I think you should jump on the M1.

Improve this answer
4
  • thanks for the quick reply. I don't see myself developing OS Apps. rather, I make heavy use of Python libraries for working with data and implementing ML algorithms. So, I'm more interested in a CUDA-equivalent support for M1 chips. All info I can find relates to graphical performances and stuff, not really the answer to my question. Do you have one? ^_^
    – andrea
    Commented Apr 7, 2022 at 17:56
  • Hi @andrea now you’re talking some good specifics. I found the WWDC videos very helpful to research this for work. If you jump to 17 minutes in this video - developer.apple.com/videos/play/wwdc2021/10038 You can see the performance of converting a python model to CoreML and the speed at which it trains and runs. Perhaps a follow on question would be best to benchmark specific model/type of ML. You probably don’t want either portable Mac for CUDA specific workloads - Mac Pro with windows is best
    – bmike ♦
    Commented Apr 7, 2022 at 19:13
  • Your comment about “workstations” is not fully supported by the latter half of the article you cite. It shows the M1’s relative performance uplift is reversed in cases with larger datasets. Seems to be an issue of overhead, which is a valid concern if one’s workload matches the one from that article, but a tad disingenuous of a comparison.
    – fyrepenguin
    Commented Apr 8, 2022 at 1:35
  • Thanks @fyrepenguin - my intent was to link CUDA to workstations and not over generalize. If you can edit the answer to be more fair, please consider that I welcome that collaboration.
    – bmike ♦
    Commented Apr 8, 2022 at 1:50

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .