PC-Bench: A Unified Benchmark for Predictive Computing in Generative AI — Standardizing Workload Traces, Computational Costs, and Forecasting Accuracy

The increasing adoption of generative artificial intelligence (AI) has introduced new challenges in evaluating computational optimization strategies under dynamic and unpredictable workloads. Predictive computing provides a potential approach to improving generative AI infrastructure by anticipating future computational demand and preparing resources before actual requests arrive. Recent research has extended this concept from proactive inference preparation to multi-layer computational orchestration and economic-aware resource allocation. However, evaluating these approaches remains difficult because existing studies may employ different request distributions, workload characteristics, hardware configurations, computational cost models, prediction horizons, and evaluation procedures. Consequently, reported improvements are not necessarily comparable, reproducible, or generalizable across different experimental environments. This paper introduces Predictive Computing Benchmark for Generative AI (PC-Bench), a unified benchmarking framework designed to standardize the evaluation of predictive computational strategies in generative AI systems. The proposed benchmark integrates three fundamental evaluation dimensions: (1) future workload and request-demand prediction, (2) computational resource and economic cost measurement, and (3) predictive accuracy and realized operational benefits. Its architecture establishes a synchronized benchmark dataset containing timestamped request streams, workload characteristics, GPU utilization traces, computational execution costs, energy consumption records, electricity pricing information, cloud GPU pricing information, and corresponding future-demand observations. Unlike conventional LLM inference benchmarks that primarily evaluate model accuracy, throughput, and latency, PC-Bench investigates whether computational demand can be anticipated sufficiently early to support economically and operationally beneficial preparation. The framework distinguishes between prediction accuracy, prediction usefulness, successful resource reuse, and realized net computational savings. It further introduces standardized prediction horizons, chronological training and evaluation partitions, workload replay protocols, cost-accounting procedures, and baseline configurations to facilitate fair comparisons between competing predictive computing methods. A unified benchmark representation and evaluation formulation are developed to characterize the relationship between predicted demand, preparation expenditure, computational resource consumption, and realized operational outcomes. The proposed benchmark includes multiple evaluation tracks covering request prediction, GPU demand forecasting, proactive inference preparation, predictive resource scheduling, economic-aware computation, and integrated predictive computational orchestration. It also incorporates uncertainty, prediction drift, resource contention, failed predictions, unused precomputation, and information freshness into the evaluation process. Rather than assuming that a single metric can adequately characterize predictive computing performance, PC-Bench proposes a multi-dimensional evaluation methodology that separately measures forecasting quality, computational efficiency, economic efficiency, service reliability, and resource utilization. The framework provides a foundation for constructing standardized datasets and reproducible experimental protocols for predictive computing research in generative AI. Its practical value will depend on the availability of sufficiently representative workload traces, reliable cost measurements, and independent empirical validation.