AI-Driven Optimization of Chiplet Architectures Using Reinforcement Learning

The increasing demand for scalable and energy-efficient processing in domains such as artificial intelligence (AI), cloud computing, and edge analytics has driven the exploration of chiplet-based architectures. These architectures enhance performance, modularity, and manufacturing yield by integrating multiple smaller dies, referred to as chiplets, within a single package, enabling heterogeneous integration and improved scalability compared to traditional monolithic systems. However, this shift introduces new challenges, including inter-chiplet communication latency, uneven thermal distribution, and inefficient workload scheduling under dynamic conditions. To address these limitations, this paper proposes a reinforcement learning (RL)-based methodology that leverages intelligent, adaptive decision-making to optimize system performance. By modeling the chiplet-based platform as a Markov Decision Process (MDP), the proposed framework allows RL agents to dynamically manage interconnect configurations, workload allocation, power consumption, and thermal balancing in real-time. Simulation results demonstrate that the RL-driven approach achieves significant improvements across multiple performance metrics, including a 52% increase in throughput, a 36% reduction in latency, a 29.2% reduction in power consumption, and a 16.3% improvement in thermal balance. These results highlight the potential of RL to enable highly efficient, scalable, and self-optimizing chiplet-based systems for next-generation high-performance computing applications.