Inhomogeneity is always the thing that keeps these architectures from being successful, in the end. It's a premature optimization that seems like the right thin

Inhomogeneity is always the thing that keeps these architectures from being successful, in the end. It's a premature optimization that seems like the right thing to do at any given moment. Then conditions change, you find out that FP32 isn't needed, and FP8 is ok... then 3 bits... then Trinary... all require new hardware. Memory that computes directly, in as simple a model as possible is required. Imagine a huge grid of LUTs that just talk to their neighbors, and I believe, you've got an approach that can work for almost any compute task.