ASIC-powered Distributed Collaborative Machine Learning

Distributed Collaborative Machine Learning (DISCO) challenge is to look for a solution to make it more carbon aware. One of the problems with ML is that training heavy models requires a lot of energy

Comment

Inspiration

DISCO is a really interesting technology, but what if the model to train is too much for a normal person's hardware? Can it also be possible to reduce carbon emissions with such a solution? https://learn.microsoft.com/en-us/azure/machine-learning/v1/how-to-deploy-fpga-web-service

What it does

Development of custom ASIC/FPGA to increase performance in the training. The training can be done on the phone of the user.

How we built it

It is possible to turn a machine learning model into an ASIC. https://towardsdatascience.com/how-to-make-your-own-deep-learning-accelerator-chip-1ff69b78ece4

Challenges we ran into

Developing an ASIC/FPGA is really expensive and hard, but with technologies such as https://circt.llvm.org/, it should be much easier than developing a VHDL solution from scratch.

Another challenge is how to deliver the hardware:

  • to the house of the person, but this will also require a return and is not eco-friendly
  • the user will have to go to the closest research centre having such hardware available.

Accomplishments that we're proud of

  • Communication from phone to PC to transfer data and train the model

What we learned

  • ASIC can be used for machine learning
  • Distributed Machine learning can improve privacy and help research

What's next for ASIC-powered Distributed Collaborative Machine Learning

  • See if there's interest in the idea
  • Partner with some Semiconductor Fabs in Europe to check the production costs

Built With

Share this project:

Updates