Unrolling thread...
We made a $900 RTX 3090 faster and more efficient than an M5 Max at running LLMs: