Can't we just use HBM like memory that acts as cache but put it on top of l3 cache like AMD is doing with it's 3d cache. That will help with latency.