Micron's memory wall chart. Compute up ~3x every two years, HBM bandwidth under 2x
Micron's memory wall chart. Compute up ~3x every two years, HBM bandwidth under 2x
Deep Dive
From Raghu Sreeramaneni's memory tutorial at hot chips 2026. top line is normalised tflops for tpu v3 through r200, bottom line is hbm2e through hbm4, both log scale, so the distance between them is a lot wider than it looks. The three boxes down the right are the fixes people are actually building.