Machine Learning in PCIe networks

Compare, benchmark, and optimize machine learning frameworks.

Machine learning is a popular topic in data science, and there exist many frameworks that can be used for training and inference of these neural networks. However, many of these frameworks are still only optimized for one machine and only the CPU architecture. In this thesis, we want to do a survey of available machine learning frameworks and find out which frameworks support which features. After the survey is complete, we want to try to minify one or more of the frameworks to run on Dolphins PCIe networks, either with socket support or by using Dolphins low-level SISCI API.


Analyze several popular machine learning frameworks (such as TensorFlow, Caffe, etc.) and find out what features such as GPU acceleration, multi-machine support the different framework support. Select one of the frameworks that are open source, and try to modify the communication system to enable multi-machine support in Dolphins PCIe networks.

Learning outcome

In-depth knowledge and understanding of optimizing a machine learning framework.


Good low-level computer systems understanding. The student should have completed, INF3151, or equivalent. IN5050 is recommended in the degree.


  • Håkon Kvale Stensland
  • Michael Riegler
  • Hugo Kohmann, Dolphin Interconnect Solutions

Collaboration partners

Dolphin Interconnect Solutions

Contact person