Sunday, November 12, 2017

TensorFlow on MX150

One can run TensorFlow on NVidia GeForce MX150 graphics card using the following setup:
  • CUDA version 8.0
  • CuDNN version 6.0 for CUDA 8.0
Timing comparison for matrix multiplication using CPU (i7-8550) (shown in orange) and GPU (MX150) (shown in blue) for increasing matrix sizes. 
 
{'/gpu:0': [0.23459196090698242, 0.5798056125640869, 0.08466935157775879, 0.0998234748840332], 
'/cpu:0': [0.7988660335540771, 4.3423473834991455, 14.469173431396484, 42.78021287918091]}

The code for doing the matrix multiplication is given here.
Posted by at
Labels: , ,

9 comments:

  1. UnknownDecember 5, 2017 at 8:56 AM

    Hi, thank you very much for this post and the provided code!
    Do you know, what CUDA compute capability level the MX150 has? There is no information on their website. Thank you!

    ReplyDelete
  2. Benjamin SautermeisterDecember 23, 2017 at 6:49 AM

    Could you post some more benchmarks using your MX150 vs CPU using examples provided by TensorFlow (MNIST, CIFAR10, RNN seq2seq)? That would be awesome! Thank you in advance! :)

    ReplyDelete
  3. VigniFJanuary 7, 2018 at 6:44 AM

    there's no official support on nvidia website for mx150 how did you do that?

    ReplyDelete
  4. Human PersonJanuary 7, 2018 at 7:30 PM

    Agreed. I'm currently trying to install Cuda 8.0.61 on windows 10 with a mx150, and the installer doesn't seem to register the mx150 gpu. Would be interested in how you accomplished getting cuda to run on an mx150.

    ReplyDelete
    Replies
    1. Human PersonJanuary 7, 2018 at 7:34 PM

      You can go ahead with the installation, but then you get a missing dll error when trying to run tensorflow.

      Delete
  5. OterLabbFebruary 22, 2018 at 5:58 AM

    I got it working on a MX150 with python 3.6 (64 bit) installed a precompiled wheel from https://github.com/fo40225/tensorflow-windows-wheel/tree/master/1.5.0/py36/GPU/cuda91cudnn7avx2

    And CUDA 9.1 and CUDNN 7.0.5

    ReplyDelete
    Replies
    1. UnknownJune 2, 2018 at 3:42 PM

      How do you use this pre-compiled wheel in the installation process?
      I'm using python 3.6, i7 cpu on W10 and MX150

      Delete
  6. UnknownOctober 11, 2018 at 1:06 AM

    Hi, i got cuda 9.0 and cudnn 9.0 wuth tensorflow, While running the keras code with tensorflow backend GPU usage stays in between 12-19%. Can you please tell me how to increase the speed.

    ReplyDelete

Subscribe to: Post Comments (Atom)