Qwen3.6-27B tests 3-GPU speed ceiling

The numbers do not look wildly off for a dense 27B model on consumer AMD hardware, but they do suggest the decode path is the bottleneck, not raw VRAM.

• –27B Q8 across 3x 7900 XTX is already a high-friction inference setup, so scaling gains will come from tuning more than from simply adding cards
• –650 t/s prompt processing is decent; the gap is that decode speed often flattens out because of split overhead, synchronization, and KV-cache behavior
• –This is a useful real-world datapoint because AMD multi-GPU llama.cpp performance is discussed far less often than CUDA/NVIDIA setups
• –The most relevant knobs are likely tensor split strategy, batch sizes, context settings, and build/driver versions rather than the model alone