User benchmarks NVIDIA CMP 170HX for local LLM inference

Testing shows that 8GB CMP 170HX mining cards can run multiple small models simultaneously or fit individual models entirely within their VRAM. The setup utilized four cards to achieve 64GB of total VRAM capacity for local inference tasks.