Benchmark notes from a tuning session on a single NVIDIA GeForce RTX 5090 running Qwen3.8-27B with two inference engines: q27 / Quasar and ninfer.
q27 (Quasar) is a specialized CUDA inference engine by signalnine targeting Qwen3.6/Qwen3.8-27B-MTP on a single RTX 5090 and related NVIDIA GPUs.
The goal was not maximum raw throughput, but finding the best operating point in terms of tokens per second per watt (tok/s/W).