Through pcie using GPU P2P with a supported AMD chipset (epyc processor) so they can talk to each other or going through your CPU and RAM. I think the latter is

Through pcie using GPU P2P with a supported AMD chipset (epyc processor) so they can talk to each other or going through your CPU and RAM. I think the latter is what people using nvidia RTX 4090s do because P2P is disabled and NVlink is unsupported