[kcb]

Problem is once you have to scale to multiple GPUs the interconnect becomes the primary bottleneck.