Final Submission

This post describes the final submission for my 2016 Google Summer of Code project GPU-accelerated Deep Neural Networks in TMVA. The work currently includes production ready implementations of accelerated training of deep neural networks on

• Multi-core CPU architectures
• CUDA GPU architectures

In addition to that, the code includes a reference implementation used for the verification and testing as well as an OpenCL implementation that is still under development.

Project Documentation

So far, I have written five blog posts describing the work progress during this summer. Additional posts will follow in the coming weeks evaluating computational and classification performance of the neural network implementation. The final presentation made to the SFT group at CERN can be found here.

Commits

The main work is contained in the three branches, one for each of the architectures:

• CPU
• CUDA
• OpenCL

Currently work is underway to combine the the implementations and prepare the integration into Root master:

• Integration