Prefill vs decode, why memory bandwidth caps tokens/s, MoE, quantization and context limits — a crash course with real measurements from a 12 GB GPU and a very patient Xeon.

The problems between the drawing board and a working vehicle: a browser-fed watchdog that flipped the OVERLOAD badge, a palette filter that ate nine cores, a camera that refused to follow orders, and a map that stalled the handheld.