Production-grade LLM inference involves complex interactions between hardware and software. Key performance indicators in this realm are shaped by the specifics of hardware, such as GPU capabilities from NVIDIA and AMD, which vary in their handling of different numeric types and bandwidths. The article emphasizes two main phases of LLM inference: prefill and decode. The prefill phase is heavily compute-bound, with high arithmetic intensity, meaning that GPUs spend more time processing than waiting for data. In contrast, the decode phase is memory-bound, where GPUs often wait for data to load before generating each token. Important metrics include Time to First Token (TTFT), which measures how quickly a model generates its first token after receiving a prompt. TTFT is closely tied to the prefill phase, while Time per Output Token (TPOT) reflects the speed of token generation during the decode phase. The article also introduces Inter Token Latency (ITL), which captures the time between consecutive tokens and can indicate performance consistency. End-to-End Latency (E2EL) measures the total time from when a user submits a prompt to when they receive the final token, factoring in network overhead, TTFT, and TPOT. Token throughput and request throughput are additional metrics that gauge the efficiency of the inference system. Token throughput counts how many tokens are generated per second, while request throughput focuses on how many user requests the system can handle over a given period. These metrics are essential for optimizing unit economics in AI applications, especially as hardware costs remain high. The article illustrates that fine-tuning these metrics can lead to substantial improvements in performance and user experience.