https://hgpu.org/?p=10170
GPU peer-to-peer techniques applied to a cluster interconnect