Nvidia already added optimized mixed-precision fused multiply-accumulate units. That’s about as good as you could ask for for specialized hardware for the kind

Nvidia already added optimized mixed-precision fused multiply-accumulate units. That’s about as good as you could ask for for specialized hardware for the kind of math current AIs use. The main bottleneck is memory bandwidth. So maybe future chips will trade off fewer or slower compute units for faster RAM.