[elmarhaussmann]

For this benchmark, NVLink and gradient reduction isn't the bottleneck. The performance scales almost perfectly linearly from one GPU to four.