This is one of the clearest breakdowns of KV cache optimization I've read. The "chef waiting for ingredients half a mile away" analogy for memory-bandwidth-bound inference is going to stick with me. Really well-structured comparison across five very different approaches!