**Introducing SPEED-Bench: A Unified and Diverse Benchmark for Speculative Decoding**

Speculative Decoding is a critical technique that accelerates large language model inference by using a smaller draft model to predict future tokens in parallel. However, current evaluation methods are often fragmented and fail to represent real-world data or serving conditions, where performance is highly dependent on the task and system load. To address this, SPEED-Bench offers a unified benchmark designed to test speculative decoding across diverse semantic domains and realistic serving regimes. It features a "Qualitative" split to measure accuracy across topics like coding and math, and a "Throughput" split to evaluate system-level speedups with large batch sizes and long inputs. This comprehensive framework allows practitioners to analyze performance in a way that more accurately reflects production environments.