P2P on RTX 5060 Ti boosts vLLM prompt processing by 25%

A savvy developer running a 4x RTX 5060 Ti 16GB rig with vLLM discovered that enabling PCIe peer-to-peer (P2P) on consumer Nvidia cards delivers a substantial performance boost. The test used a Qwen3.6-27B-FP8 model with tensor parallelism on an EPYC server with 8-channel memory. With P2P disabled, prompt processing (pp2048) hit 1,649 tokens/s; after enabling P2P via patched drivers and Resizable BAR (ReBAR), it jumped to 2,305 tokens/s — roughly 25% faster. Token generation also improved at 8K context, rising from 109 to 123 tokens/s.

The trick requires hardware with ReBAR support and installing patched kernel modules from the GitHub repo aikitoria/open-gpu-kernel-modules. The gains come from allowing GPUs to communicate directly over PCIe, bypassing CPU RAM bottlenecks. Even on high-bandwidth systems like this EPYC server, P2P cut time-to-first-token substantially, from 1,242ms to 913ms at pp2048. For developers building cost-effective local inference rigs, this is a free performance unlock that makes consumer cards behave closer to datacenter parts in multi-GPU setups.