how can i tell if GPU access is available?

User 2607 | 11/14/2015, 6:18:01 PM

before i start training deep features i want to know if i have everything correctly set to use the GPU. is there function to inspect this?

Comments

User 2600 | 11/14/2015, 10:22:54 PM

I would also like to be able to sanity check this. Really there is no configuration to use this feature? How can I be reassured that GraphLab is actually offloading computations to the GPU cards, besides running tests between a non-GPU enabled install and after the upgrade, which to my mind defeats the purpose of that (by the way totally awesome) simplicity? And what if there's two NVIDIA GPUs attached to the same machine? I feel like I missed the high-level overview of how GraphLab provides GPU acceleration automagically - a pointer to that material would be of great interest.

Thanks!


User 940 | 11/17/2015, 8:05:58 AM

Hi @clyde and @dylanht ,

Currently, the best way to check if the GPU card is the following:

`python data = graphlab.SFrame('http://s3.amazonaws.com/dato-datasets/mnist/sframe/train')

m = graphlab.neuralnet_classifier.create(data, target='label') `

Now in the output, before training progress, if you see:

PROGRESS: Creating neuralnet using cpu

that means the GPU is NOT being used. Otherwise, if you have GPU's they should be listed. If you have multiple GPU's, they should all be leveraged.

Currently, neural nets are the only toolkit accelerated by GPU cards, and fundamentally it's large matrix multiplications that are being offloaded onto the card.

Honestly, this is a bit of a clunky way to check if the GPU is being utilized. I'll make a note to write a utility function to check this.

Hope this helps!

Cheers! -Piotr


User 2600 | 11/18/2015, 7:18:49 AM

Thanks @piotr, good to know!