search

InonS / Open-GPGPU-ANN

Open Source GPGPU support for Artificial Neural Networks
1 stars 0 forks source link

Any luck on OpenCL? #1

Open Dexdev08 opened 7 years ago

Dexdev08 commented 7 years ago

I actually wanted to utilize my Intel HD5000 GPU on my laptop (although the RAM is a bit small for this). Any luck on your side?

I have tried Hugh's libraries - I'm a bit of a n00b so for keras, i'm using the newer version.

InonS commented 7 years ago

At this point I'm not even sure what I thought was successful even was...

Please see my conversation with Hugh https://github.com/hughperkins/tf-coriander/issues/60. TL;DR it ends with both of us giving up.

https://github.com/hughperkins/tf-coriander/issues/60

Dexdev08 commented 7 years ago

I see! I really wish opencl was more developed in deep learning. I wanted to use existing hardware.

At this point, i am decided to bite the bullet and just setup my own deep learning rig with an nvidia gpu.

Dexdev08 commented 7 years ago

My experience is that my gpu in a laptop, even if i get tf corainder running on toy problems, will not be able to handle any practical problems. Note that the ram is just 384mb for the gpu.

hughperkins commented 7 years ago

Yes, you'd never train on a laptop gpu. You might use it for development, but any library in mainstream usage, notably Caffe, Torch, pyTorch, Tensorflow, ... will let you develop on your cpu, and then train later on a GPU, once you've got it working on the cpu anyway.

Personally, I dont have access to any discrete AMD GPUs. There are none in the Cloud anywhere, that I've seen. That limits the extent to which I can use myself any OpenCL products I create :)

InonS commented 7 years ago

I was hoping to develop on my laptop AMD GPU, thinking it would be faster than on my CPU. An iterative development process is essentially. If the production model (implemrnted in OpenCL) could run on nVidia GPUs, that would complete the cycle.