Goals

  • Finish TVMCon slides

I think it’s finally time to do some kind of project involving memory and accelerators. It’s possible that we don’t have much to do; it could be that Dynamic Tensor Rematerialization will work for accelerators just as well as it works for GPUs. And why not? It would just require some custom interfacing with the accelerator’s on-chip memory, and probably tuning of the heuristic.