Deep Neural Network Training on GPU

[francescobrozzu]

Hello everyone!

I'm currently testing out the Intel oneAPI AI toolkit on Devcloud. My current objective is to train and evaluate a custom Deep Neural network model for image segmentation. By checking the samples I've notices that this can be achieved by using Tensorflow with the oneDNN optimizations enabled. After checking whether they were being used by calling "isMKLEnabled()" I've started the training procedure, however, I've noticed that the performance seems far inferior with respect to other GPUs.

Looking thoroughly at the documentation for oneDNN and Intel optimized TensorFlow it seems that TensorFlow is not really running on the Intel GPU, but rather on the CPU with some specific optimization. Therefore my question is: is there a toolkit (not necessarily Tensorflow) that allows to train and evaluate a custom deep neural network on an Intel GPU using oneDNN? Or alternatively, is it possible to use the oneDNN primitives to build a custom application that can train the neural network using the GPU?

Best regards,

Francesco Brozzu

[RaeesaM_Intel · Accepted solution]

Hi Francesco Brozzu,

Thank you for posting in Intel Forum.

Your findings are correct. Tensorflow available with AI analytics toolkit has been directly optimized for Intel® architecture using the primitives of Intel® oneAPI Deep Neural Network Library (oneDNN) to maximize performance. This package provides the latest TensorFlow binary version compiled with CPU enabled settings (--config=mkl).

If you want to train and evaluate a custom Deep Neural network model, We have Intel® oneAPI Deep Neural Network Library included in Intel® oneAPI DL Framework Developer Toolkit that helps you develop fast neural networks on Intel® CPUs and GPUs with performance-optimized building blocks. 

You can refer the following documentation of DL kit for more information :

https://software.intel.com/content/www/us/en/develop/tools/oneapi/dl-framework-developer-toolkit.html

Get back to us incase of any further clarifications. We would be happy to help you.

Regards,

Raeesa

[RaeesaM_Intel]

Hi,

Thank you for accepting the solution.

This thread will be no longer supported by Intel. Please raise a new thread if you have any further queries.

Regards,

Raeesa