Running on GPUs other than CUDA-enabled

Deep Learning
1

Hi everyone!

Not sure whether this has been discussed previously, but I thought I would throw it here since after a few searches it does not appear to me this topic has ever been raised (should I be wrong please point me to the relevant section).

I have recently purchased an Nvidia-equipped laptop specifically to try out a few DL models locally. It turns out I should have done my homeworks better, as the GPU it came with is among those which are not CUDA-enabled… Now, from quick searches on the internet it seems Theano may still run using non CUDA libraries (Opencl?) and I wonder whether anybody has found her/himself in the same situation and if a workaround exists.

To summarize: can I run Theano on GPU, even though the GPU is not CUDA-enabled? If so, what need to be done?

Thanks in advance!
Luca

1 Like
2

I think you cannot achieve much since Theano (or Tensorflow) forwards many things to cuDNN. Have you set environment variable THEANO_FLAGS to device=opencl? It should be enough to use opencl.

3

Reith,

thank you for your reply. I’ll give it a try with opencl, although as you say I should probably manage expectations.