P2P Disabled Cuts Qwen3.8-27B Throughput on Radeon R9700s

An author benchmarked tensor-parallel LLM serving of Qwen3.8-27B across four AMD Radeon AI PRO R9700s with GPU peer-to-peer (P2P) enabled and disabled. Turning P2P off substantially reduced prefill token rates by up to 55 percent, dropped decode throughput, more than tripled median time to first token under load, and increased energy consumption per output token by 59 percent. The results demonstrate that direct card-to-card communication via P2P is crucial for efficient tensor parallel inference performance. Open article →