153 tok/s on 1x AMD Radeon R9700 running Qwen3.8 27b NVFP4, 470 tok/s @ 8 conc requests, Prefill @ 3,619 tok/s

Optimizations for single AMD Radeon R9700 cards have significantly improved performance, doubling speeds across various tasks. Running Qwen3.8 27b NVFP4, the card now achieves 153 tok/s for JSON processing and 141.7 tok/s for summarization. Prefill speeds reach 3,619 tok/s at a depth of 16000, and with 8 concurrent requests, the card delivers 471 tok/s.

This report details how recent optimizations have doubled the performance of single AMD R9700 cards, unlike previous benchmarks that showed lower speeds.