2x 5060 Ti Qwen3.6 Benchmark

Qwen3.6-27B and Qwen3.6-35B-A3B were measured under multiple configurations on a 2-card RTX 5060 Ti 16GB setup.

• The measurement tool was llama-benchy 0.3.5, using the conditions --pp 4096 --tg 128 --depth 0 --runs 3 --latency-mode generation --no-cache.
• For Qwen3.6-27B, vLLM's NVFP4-MTP, TP2-PP1, no spec combination had the highest PP throughput at PP 1963 t/s, TG 38.4 t/s, TTFT 2182ms.
• For the same model, vLLM's Lorbus/Intel AutoRound series was generally around PP 1044~1088 t/s, TG 40.2~46.9 t/s, TTFT 3792~4008ms.
• ik-llama.cpp showed PP 1450 t/s, TG 28.4 t/s, TTFT 2945ms with IQ4_XS layer + q8_0 KV, while tensor + f16 KV showed lower PP and higher TTFT.
• For Qwen3.6-35B-A3B, vLLM's NVFP4, TP2-PP1, no spec combination was best at PP 6259 t/s, TG 116.5 t/s, TTFT 753ms.
• On the other hand, vLLM's DFlash n=15 dropped to TG 38.9 t/s, and ik-llama.cpp recorded TG 108.9 t/s with Unsloth Q4_K_XL layer + q8_0 KV.

The author stated that speculative decoding effectively failed due to PCIe bandwidth limitations, and mentioned plans to re-measure with larger pp/tg settings.