Goals

  1. Blog-ify LATTE thoughts
  2. Make project proposal for UCSB people
  3. Draft eval section for Arxiv
  4. Get as far as I can in getting Chien Yu’s code working
  5. Make Python notebook end-to-end

A core research idea coming from my work so far is that data movement/data access patterns should be first-class citizens, at least in machine learning. Often, how you orchestrate the data is more important for performance than how you compute over it. This applies at every scale, from the system-level concern of transferring between main memory and GPU memory, to the accelerator design concern of how you feed a systolic array.

I didn’t do well on my goals, but certainly worked hard. I finally let myself put some time into making a web demo for Glenside, which is live (though not actually doing anything) here. This was a guilty pleasure; at least, that’s what I was telling myself. I wanted to do it, so I didn’t let myself do it. I do this frequently; I tell myself the thing I want is not the thing I should do. There are points when that’s true, certainly, but I think I take it too far sometimes.