How would you architect a system to maximize data throughput between multiple GPUs and system memory at Advanced Micro Devices?
