WWW: What, When, Where to Compute-in-Memory
The episode discusses a paper titled "WWW: What, When, Where to Compute-in-Memory for Efficient Matrix Multiplication during Machine Learning Inference." The authors provide an analytical framework addressing what compute-in-memory primitive to use, when it benefits specific workloads like matrix multiplication shapes, and where in the memory hierarchy it should be integrated. The hosts conclude that capacity in on-chip memory is key, but compute-in-memory is not a universal solution.