I think that silicon and data compression are well-optimized domains already.

However, Neural Nets are provably horrible w.r.t resource efficiency, as demonstrated by recent advances like EfficientNet, which managed to decrease the number of parameters drastically and improve top 1% performance as well.