5 ms·

Can anyone comment on whether this will be useful for training neural networks?

kogir 9y ago
Yes, it will allow you to use a much more powerful graphics card, as long as your training isn't limited by bus bandwidth.
- malcolmgreaves 9y ago
  Only if the software you have supports it. For example, Tensorflow on OS X doesn't use GPUs since version 1.2
minimaxir 9y ago
No. Officially, this doesn't support NVIDIA cards yet, which are a requirement for any modern DL framework. There are beta NVIDIA drivers (http://www.nvidia.com/download/driverResults.aspx/131833/en-us http://www.nvidia.com/download/driverResults.aspx/131833/en-...) for macOS, but your mileage will vary.
- samcat116 9y ago
  These aren’t beta drivers. They’ve been released for quite some time and are updated with (almost) every macOS release. However I’d say they’re pretty barebones as Nvidia doesn’t seem to care too much about the Mac at this point.
  - minimaxir 9y ago
    The notes say "Includes BETA support for iMac and MacBook Pro systems with NVIDIA graphics." (specifically, Pascal drivers) The reason for that is Apple hasn't shipped a iMac/MacBook Pro with an NVIDIA card in years, which is especially annoying for DL work.
    - michaelgreen 9y ago
      In practice they work just fine, it's just very annoying to initially set them up and you have to do some patching to get the latest version of tensorflow working with it. I use one every day and I haven't run into any problems while running models with both pytorch and tensorflow.
mkagenius 9y ago
There is a project working for this: https://github.com/plaidml/plaidml https://github.com/plaidml/plaidml
bitL 9y ago
IMO it's more cost effective and stable just to buy a Ryzen 2400G system (the same cost as a good eGPU case), put in Linux and use it for training at a full speed. TB3 has 4x lower throughput than PCIe, so if you need to transfer a lot of data (any meaningful non-toy model these days), it's better to stick with internal GPU. Not mentioning Linux is 1st class/preferred platform for many frameworks like TensorFlow etc.
- walshemj 9y ago
  I say for thease sort of application why bother with MAC at all just go with Ryzen or wait for Ryzen2 out this month or go the full Threadripper or Eypc / Rome
michaelgreen 9y ago
So I've been using an eGPU for about 3 months now and it is amazing. This isn't officially supported so you end up having to do a lot of work arounds to get things like tensorflow/pytorch working. It doesn't let you hot plug your eGPU into your computer (you have to restart for it to work) but for officially supported eGPUs you're now able to do that. I've trained various models using a Titan XP and it's so awesome to be able to maintain the portability of your laptop and still get all of that power. A large benefit is also not having to move training data around between servers and other machines if you have it on an external drive or just on your laptop. After you get everything up and running there isn't that much maintenance or anything you have to do regularly to keep it working. The speedup is incredible and it was definitely worth it for me personally, however it's not a walk in the park to get it set up initially.
- dmix 9y ago
  Neat. Have you tried it with any games?
  - michaelgreen 9y ago
    I've tried it with an external monitor which works great but I haven't tried it with any games. From looking at the eGPU forums it seems like people aren't running into that many problems with it.

Yes, it will allow you to use a much more powerful graphics card, as long as your training isn't limited by bus bandwidth.

malcolmgreaves 9y ago
Only if the software you have supports it. For example, Tensorflow on OS X doesn't use GPUs since version 1.2

minimaxir 9y ago

No. Officially, this doesn't support NVIDIA cards yet, which are a requirement for any modern DL framework. There are beta NVIDIA drivers (http://www.nvidia.com/download/driverResults.aspx/131833/en-us http://www.nvidia.com/download/driverResults.aspx/131833/en-...) for macOS, but your mileage will vary.

samcat116 9y ago
These aren’t beta drivers. They’ve been released for quite some time and are updated with (almost) every macOS release. However I’d say they’re pretty barebones as Nvidia doesn’t seem to care too much about the Mac at this point.
- minimaxir 9y ago
  The notes say "Includes BETA support for iMac and MacBook Pro systems with NVIDIA graphics." (specifically, Pascal drivers) The reason for that is Apple hasn't shipped a iMac/MacBook Pro with an NVIDIA card in years, which is especially annoying for DL work.
  - michaelgreen 9y ago
    In practice they work just fine, it's just very annoying to initially set them up and you have to do some patching to get the latest version of tensorflow working with it. I use one every day and I haven't run into any problems while running models with both pytorch and tensorflow.

mkagenius 9y ago

There is a project working for this: https://github.com/plaidml/plaidml https://github.com/plaidml/plaidml

bitL 9y ago

IMO it's more cost effective and stable just to buy a Ryzen 2400G system (the same cost as a good eGPU case), put in Linux and use it for training at a full speed. TB3 has 4x lower throughput than PCIe, so if you need to transfer a lot of data (any meaningful non-toy model these days), it's better to stick with internal GPU. Not mentioning Linux is 1st class/preferred platform for many frameworks like TensorFlow etc.

walshemj 9y ago
I say for thease sort of application why bother with MAC at all just go with Ryzen or wait for Ryzen2 out this month or go the full Threadripper or Eypc / Rome

michaelgreen 9y ago

So I've been using an eGPU for about 3 months now and it is amazing. This isn't officially supported so you end up having to do a lot of work arounds to get things like tensorflow/pytorch working. It doesn't let you hot plug your eGPU into your computer (you have to restart for it to work) but for officially supported eGPUs you're now able to do that. I've trained various models using a Titan XP and it's so awesome to be able to maintain the portability of your laptop and still get all of that power. A large benefit is also not having to move training data around between servers and other machines if you have it on an external drive or just on your laptop. After you get everything up and running there isn't that much maintenance or anything you have to do regularly to keep it working. The speedup is incredible and it was definitely worth it for me personally, however it's not a walk in the park to get it set up initially.

dmix 9y ago
Neat. Have you tried it with any games?
- michaelgreen 9y ago
  I've tried it with an external monitor which works great but I haven't tried it with any games. From looking at the eGPU forums it seems like people aren't running into that many problems with it.