This is a way better explanation of Dynamic Caching than what most techtubers have vaguely mumbled about it with no clear understanding.

It said on-chip memory right on their slide which seemed to have been lost on most, which is SRAM and not DRAM, the layers of GPU caches/tile memory and fitting more as it’s needed in there which maximizes GPU hit rates and utilization. Most people explained it in some hand wavey nature about Unified Memory and using less RAM.