4 ms·Problem is once you have to scale to multiple GPUs the interconnect becomes the primary bottleneck.by kcb 2y agoProblem is once you have to scale to multiple GPUs the interconnect becomes the primary bottleneck.