I'm building convoluted neural networks with TensorFlow. I've got the latest Mac Pro, use all of my cores, and yet it still takes me several hours to train a single network. I need to do grid search on many different tuning parameters, so this is not ideal.

This is for a personal project, so I want to save money by using the cloud. I have used AWS extensively, and I like the idea of using their spot market, however I've heard that 1) Amazon's GPU instances aren't compatible out of the box with TensorFlow (AWS's GPUs have CUDA 3.0 and TensorFlow requires 3.5) and 2) Amazon's GPUs are a bit slow.

Does Microsoft's Azure or Google's Google Compute Engine have easy-to-use GPUs when using TensorFlow? How do their GPUs compare to those of Amazon in terms of the relevant neural-net-training tech specs, prices, and ease of use? Are there smaller cloud providers that you would recommend instead?

Also, roughly how much faster can I expect training to take on one of these virtual GPUs compared to a top of the line Mac Pro?

  • 1
    $\begingroup$ Can you give details on : 1. Whether you are using GPU on your Mac currently 2. What's the desired training time that you want? . A couple of hours is not unusual to train typical deep networks. $\endgroup$
    – wabbit
    Jul 12 '16 at 14:12
  • $\begingroup$ The TensorFlow binaries for Mac currently only support CPU, not GPU (tensorflow.org/versions/r0.9/get_started/os_setup.html). The desired time is as fast as possible. I accept that it will always probably take hours, but I'd like to reduce the time as much as possible $\endgroup$
    – Ryan Zotti
    Jul 12 '16 at 14:24
  • $\begingroup$ Maybe you can check if you can compile tensorflow for your Mac gpu. Compiling is quite easy as per my experience compiling on AWS for the g2.8xlarge instance $\endgroup$
    – wabbit
    Jul 12 '16 at 14:35
  • $\begingroup$ Ok, I'll check that out $\endgroup$
    – Ryan Zotti
    Jul 12 '16 at 15:14
  • $\begingroup$ See github.com/tensorflow/tensorflow/issues/25 for how to compile for AWS $\endgroup$
    – wabbit
    Jul 13 '16 at 10:33

Microsoft announced a couple of weeks ago virtual machines in Azure with GPUs. They use K-80 NVIDIA cards. The biggest machine has 4 GPUs and 224 GB of ram.

Good to play with deep learning :-)


Google does not currently have any public GPU or TPU offerings, but they might be able to help if you work for a large company. Contact them directly.

Source: Google reps, just today.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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