[Const-me]

> How did we overcome the HBM <-> SRAM bottleneck?

Because every number we load from the model through that bottleneck gets reused, to compute different requests within the batch.