Really cool. I'm very much still learning about this stuff. Sounds like this inter-GPU communication is a feature of special hardware (not consumer GPUs).
GPU interconnect speeds are a big bottleneck today for GPU's in AI applications. Data can't move between them fast enough.