GoKawiilViiwork tests KV-cache sharing across two RTX 5090 hosts over 1GbE vs 10GbE
A benchmark compared viiwork's kv_share feature on two RTX 5090 machines connected via 1 GbE, a direct 10 GbE cable running Tailscale, and the same 10 GbE cable sending kv_share traffic directly. The test ran a coding-agent workload building a todo app, measuring 12-session and 16-session runs over 90 minutes to see how network speed affects whether a session's cached prompt reaches a second host before it's needed.
GoKawiil's interpretation of the reporting above, not reported fact.
KV-cache sharing avoids expensive prefill recomputation when a coding-agent session spills to a less busy host, so the speed of the link between machines could determine whether that cache arrives in time to actually help. The results suggest multi-GPU-host setups for LLM agents may need faster networking than typical LAN gear to realize the full benefit of cache sharing, which matters for anyone scaling agent workloads across cheaper consumer GPUs rather than a single large server.
- Viiwork's kv_share copies a session's KV cache to the next likely host after each turn to avoid cold-start prefill delays.
- The test compared 1 GbE, 10 GbE with Tailscale, and 10 GbE with direct kv_share traffic across two RTX 5090 machines.
- Performance was measured under 12-session and 16-session coding-agent workloads over 90-minute runs, each host running four model slots.
NVIDIA GeForce RTX 5090 GPU — This article's whole experiment runs on a pair of RTX 5090s, making it the natural centerpiece for anyone building a multi-GPU inference or coding-agent rig. Its large VRAM and compute headroom make it well suited to hosting hefty KV caches and running local LLM workloads. Pair it with a 10GbE NIC and you're set up to replicate the exact setup described here.</br>
See NVIDIA GeForce RTX 5090 GPU on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: viiwork.pulu.app — Jani Tarvainen, 2026-10-27
Published there as: “Two RTX 5090 hosts sharing a KV cache over Ethernet”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.