Two RTX 5090 hosts sharing a KV cache over Ethernet
Summary
This viiwork post investigates KV-cache sharing between two RTX 5090 hosts to reduce recomputation during coding-agent sessions. It compares 1 GbE, 10 GbE via Tailscale, and direct 10 GbE connections, detailing how shared caches affect warm starts, transfer times, and overall responsiveness. The findings show 1 GbE is workable, while faster links mainly improve replica transfer times; most gains come from engine import timing and cache correctness rather than bandwidth alone.