batch-1-decode

This paper investigates the performance gap in batch-1 LLM decode for physical AI systems, finding that faster memory bandwidth does not proportionally reduce latency due to launch overheads, and that quantization efficiency varies significantly across hardware.