Hey folks,
Has anyone tried (or is thinking about) putting together an external GPU kit so you won’t have to use AWS for training models?
I know it’s possible but haven’t dug deep enough myself yet.
You’d be better off buying a cheap desktop and putting a GPU in it, rather than adding an external GPU to a macbook. The external GPU approach is here: https://bizon-tech.com/us/bizonbox2s-egpu.html/ . As you see, MacOS doesn’t support recent GPUs - and you’d definitely want a recent GPU, because they are much better for deep learning.
For the same price you can buy a powerful used desktop, eg: http://www.used-pcs.com/dell-xps-8700-twr-core-i7-4770-cpu-3ghz-16gb-ram-160gb-hd-cosmetic-grade-b-22840151/ . Install Linux on it, and add a recent GPU (eg GTX 1070 or 1080).
2 Likes
It’s possible now with Razer Core, however, I found it much cheaper to just go with an EC2 instance on Amazon (p2.xlarge has a Tesla K80 for $0.90 an hour…).
That has changed now: pascal drivers are there http://www.nvidia.com/download/driverResults.aspx/117854/en-us I am a laptop user as well and would love to have a quick GPU for deep-learing available locally like the 1080-TI.
Some folks mentioned that the connection throughput of the thunderbolt 3 cables compared to putting the GPU into a normal computer is inferior. I wonder if this really is an argument against e-GPUs for deep-learing as the data is transferred only once and then the number crunching starts.