Matlab limits GPU utilization?

I am trying to train a network to find knee joints in x-rays on a new PC with an Nvidia 1080ti running trainFasterRCNNObjectDetector using R2018a. The GPU utilization reported by the windows 10 task manager never exceeds 50%. Is Matlab trainFasterRCNNObjectDetector configured to limit GPU utilization? It takes so long to train that it would be great if it would use closer to 100% of the GPU. I have tried alexnet and am training with vgg16 currently.

[Rik]

You should check if this is the case as well for R2017b, because if I recall correctly, the pre-release license agreement forbids this kind of public communications.

Answer 1

[John Hipp]

The same happens when running on R2017b. I did not realize that about the license. Sorry.

[Steven Lord]

The answers and comments by Birju and Joss in this Answer may be applicable to your training code as well.

Answer 2

[John Hipp]

Thank you. That answer you referred to sheds some light. I was working on a 4 x TitanXP machine running Ubuntu. Training a VGG-19 based classification network it used all 4 GPUs to 100%. That turned out to be a problem as the power supply was likely not powerful enough and the machine would crash shortly after starting training. I had to run nvidia-smi to limit power to the GPUs and then training worked GREAT and used all 4 GPUs. However, trainfasterRCNN would only use one GPU, and a Mathworks consultant suggested moving to windows so I did. I will wait until the software is updated and hopefully this will work better some day. I hope it is OK to comment that I have had greater success with trainFasterRCNNObjectDetector with the prerelease version.

Answer 3

[Joss Knight]

I'm not sure I understand why it was recommended to you to move to Windows - can you explain?

The way the Task Manager defines utilization, I believe I would have expected it to be higher. What is the size of the parallel pool that opens when you start training? It will probably help to increase the number of workers in that pool.