[Warning: Complicated] - Codesigning AI Hardware and Software

This is a recording of a presentation I gave to a student design team at the University of Waterloo. The topic is how to make computer chips like GPUs and TPUs that can run AI algorithms efficiently.

Obviously, this is a complex topic. If you haven't taken a first year university course in linear algebra and circuits, you'll be confused. If you're not familiar with the math behind AI algorithms (like convolution operations or matrix matrix multiplication), you'll also be confused.

Anyways, if you're interested in learning more about the details, here are some notes from the textbook "Efficient Processing of Deep Neural Networks": https://docs.google.com/document/d/17...

Chapters:
0:00 Intro and pre-requisites
0:57 Efficient hardware goals
4:56 Mathematical optimisation - Toeplitz matrices
14:17 Software optimisation - operator fusion
19:10 Architectural optimisation - network on chip
26:54 Hints about gate-level optimisation
29:50 Wrapping up