After some testing, I still cann't train with two GPU in Windows 10.

But I can train with Mathematica 11.3 in Ubuntu 16.04 with two graphics card GTX 1080

The problem is, If I use two card to train, the model is not so trainable as we can see the loss curve in the following pictures. Not the reason of BatchSize's setting, I've tested the BatchSize Value From 50 to 250 in the case of TargetDevice->{"GPU",All}.

Is this a bug?

Trainable use One GPU


You may try the MNIST example

NetTrain[NetModel["LeNet"], trainingData, ValidationSet -> testData, TargetDevice->{"GPU",All}]
  • $\begingroup$ According to this answer, multiple GPU training isn't currently functional in Mathematica 11.3. I'd guess 12 is coming soon, hopefully multi-GPU will be in that release. $\endgroup$
    – Carl Lange
    Commented Sep 15, 2018 at 20:18


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.