What NPU Tops Doesn't Tell You: How Memory Bounds Local LLM Performance | Dark Hacker News