Cheapest GPU that does the job


(Terence Lewis) #1

The author at http://timdettmers.com/2014/08/14/which-gpu-for-deep-learning/ recommends the Nvidia GTX 1060 as the best buy if you want to do serious work but are strapped for cash. Has anyone used this for neural network machine translation?
Thanks.
Terence


(Etienne Monneret) #2

I’m having fun starting experimenting with the GTX 1070. But, as you write “serious work”, you certainly mean “large data”. In fact, it comes very quickly to play with the limit of 8Gb of RAM on the GTX 1070… Like me, you will certainly succeed for a lot of things with the 6Gb of the GTX 1060, but also quickly encounter some frustrations…
:wink:


(Guillaume Klein) #3

It seems to be a good entry-level GPU for deep learning.

Be careful about the memory size you get though. We usually use and recommend 8GB GPUs but you can already do a lot with the 6GB version of the 1060.

4GB or less is starting to be very limiting in this type of application.


(Terence Lewis) #4

Hi, May I contact you off forum about the 8GB GTX 1070 which I have installed? My email is lingobloke@yahoo.es.
Regards,
Terence


(Terence Lewis) #5

For anyone interested: I installed the GeForce GTX 1070 (8GB). Training from the sample data took 8 hours with the CPU. Training with the same data completed in around 20 minutes with the GPU. This cost me £399 to supply and install.
Terence


(Terence Lewis) #6

I’ve now had some lengthier training sessions with my GTX 1070 (8GB). Training a model from 3.2 million sentences took 25 hours. Another model from 1.2 million sentences took just on 17 hours (surprisingly in view of the first training as all the other parameters were the same). A colleague commented that the price of £399 represents value for money compared with training on AWS.


(Terence Lewis) #7

For those interested in speed of training: using the GTX 1070 (8GB) training a model with 10 million sentences took 5.5 days.