“The performance gains of new GPU generations are often limited by memory bandwidth, which typically does not improve by more than 2x, creating a bottleneck that prevents full utilization of increased on-chip FLOPs.”