CPU-Aware Predictive Computing for Generative AI: Anticipating Future CPU Workloads and Preparing Processing Resources Before Demand Materializes

The increasing deployment of generative artificial intelligence (AI) systems has created computational infrastructures in which central processing units (CPUs) and graphics processing units (GPUs) perform increasingly interdependent operations. Although GPU acceleration receives substantial attention in large language model (LLM) inference optimization, CPU resources remain essential for request processing, tokenization, prompt preparation, data movement, retrieval operations, scheduling, sampling, post-processing, memory management, and inference orchestration. Consequently, fluctuations in incoming AI requests may produce substantial CPU workload variations that affect overall inference throughput, response latency, and infrastructure utilization. Existing resource management approaches predominantly allocate CPU resources in response to observed workload changes, while conventional predictive computing research has primarily emphasized future information demand, GPU resource forecasting, and proactive inference preparation. This paper introduces CPU-Aware Predictive Computing for Generative AI (CPU-PCGA), a conceptual framework that investigates whether future CPU workloads can be anticipated and corresponding processing resources prepared before actual computational demand materializes. The proposed framework integrates CPU workload forecasting, request-stage workload characterization, predictive core allocation, thread-pool preparation, CPU memory and processing capacity management, workload-aware scheduling, and adaptive resource provisioning. It considers the heterogeneous CPU requirements associated with different generative AI operations, recognizing that an increase in request frequency does not necessarily produce a proportional increase in CPU utilization. The framework further investigates the relationship between predicted CPU demand and GPU execution requirements, including the possibility that insufficient CPU processing capacity may delay GPU task submission and reduce effective accelerator utilization. A unified optimization formulation is developed to determine when and how CPU resources should be reserved, activated, redistributed, or released according to predicted workload requirements, resource availability, preparation expenditure, and service-level constraints. The framework also introduces CPU readiness as a predictive resource management objective, distinguishing between forecasting future CPU utilization and ensuring that sufficient CPU processing capacity is operationally available when demand arrives. An experimental methodology is established to evaluate forecasting accuracy, CPU resource readiness, request-processing latency, CPU-GPU coordination, infrastructure utilization, energy consumption, and total computational expenditure. Particular attention is given to the distinction between aggregate CPU utilization prediction and stage-specific workload prediction, as well as the costs associated with unnecessary resource reservation and premature activation. The proposed research does not assume that proactive CPU preparation is universally beneficial; its effectiveness depends on workload predictability, CPU resource elasticity, preparation overhead, workload characteristics, and coordination with GPU execution. By extending predictive computing toward CPU resource preparation, this research investigates an additional dimension of predictive computational orchestration for increasingly heterogeneous generative AI infrastructures.