Memory move is the bottleneck these days, thus the expensive HBM, Nvidia's design is also memory-optimized since it's the true bottleneck chip wise and system wise.