What is the best GPU for deep learning currently available on the market?

I've heard that Titan X Pascal from NVidia might be the most powerful GPU available at the moment, but would be interesting to learn about other options.

And additionally - what might be considered a good machine for data-intensive computing, in terms of processor parameters, RAM etc. ?


I would recommend you read this article carefully: http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/


  • The most important feature for judging deep learning performance is memory bandwidth. GTX 1080 Ti's memory bandwidth is 484 GB/s and TITAN X is 480 GB/s. Read this Quora answer to understand why memory bandwidth is the most important feature better: https://www.quora.com/Why-are-GPUs-well-suited-to-deep-learning.

  • Tim has also given a "theoretical" rough comparison between GPUs based on their tech specs. You can also see that in his comparison 1080 Ti and TITAN X are roughly at the same performance level.

  • Use 2 1080 Ti instead of a TITAN X since GTX 1080 Ti is definitely cheaper.

  • I've personally evaluated the performance of 1080 Ti vs. TITAN X and they are roughly the same.


Look at the CUDA compute capability. They are a mixture of hardware and software features a GPU has (see guide).

I benchmarked the GTX 1070, Titan Black, GTX 970, GTX 980, GTX 980Ti. The numbers can be found in my masters thesis (Table 5.3 and Table 5.16), but the gist is:

  • GTX 1070 is by far the fastest
  • GTX 980 and 980 Ti are pretty much the same, but only ~half as fast as the GTX 1070,
  • GTX 970 needs 1.5 the time of the 980

You should look at how much memory you need. This is sometimes a limiting factor.

  • $\begingroup$ Might be worth giving the comparisons as relative scores, because "50% slower" is ambiguous - it could mean A = 100, B = 50, or it could mean A = 100, B = 66. $\endgroup$ – Neil Slater Apr 29 '17 at 16:27
  • $\begingroup$ @NeilSlater I adjusted the phrasing. In my masters thesis, I have the numbers in milliseconds for inference / seconds for training per epoch. $\endgroup$ – Martin Thoma Apr 29 '17 at 16:36

Not the answer you're looking for? Browse other questions tagged or ask your own question.