Two RTX 5090 hosts share a KV cache over Ethernet, and 1 GbE is enough
Two RTX 5090 hosts sharing a KV cache over Ethernet
A coding-agent workload ran on two RTX 5090 machines while the link between them changed: 1 GbE through a switch, 10 GbE over a direct cable with Tailscale, and that cable with kv_share sent directly. A 247 MiB cache delta took 3.24 seconds at 1 GbE, 0.81 seconds over Tailscale, and 0.55 seconds direct. Warm host switches rose from 93.8% to 96.8%, but the 2.8-second import and an import-once cache bug stayed the real bottlenecks.
For these two hosts and this workload, 1 GbE was workable: average kv_share traffic was roughly 320–450 Mb/s in each direction, and most switches were warm.