The reason it takes so much energy is that memory is separate from compute in most architecture s, so you end up getting an expensive bottleneck that is almost

The reason it takes so much energy is that memory is separate from compute in most architecture s, so you end up getting an expensive bottleneck that is almost always the limit to performance on a given task. Most of the silicon area and power go into trying to minimize the hit of that bottleneck to raw performance. Alternatives exist, they're radically different to program, but I strongly believe offer very cheap Petaflops for the masses.