Reinforcement Learning and Convolutional Neural Networks for Generalized and Adaptive Multi-Objective Mapping of DNN Workloads in Networks-on-Chip

As transistor technology advances and miniaturizes, single chips or multiple chiplets can integrate hundreds to thousands of cores. Networks-on-Chip (NoCs) have become the standard on-chip communication fabric for many-core systems, offering scalability, parallelism, and power-efficiency advantages over traditional buses. Due to these NoC properties, computations and communications of different deep neural networks (DNNs) layers can be performed efficiently on NoCs. However, traditional mapping strategies may not be suitable for running DNNs due to the different types of communication patterns, such as communication between DNN layers and one-to-many and many-to-one communications in fully connected and convolution layers. Due to varying communication patterns, it is necessary to map computations for different layers of a DNN in a manner that mitigates the communication bottleneck within the NoC. The proposed work presents a generalized framework that includes reinforcement learning (RL) with convolutional neural network (CNN) for scalable and adaptive mapping of DNN workloads to various NoC architectures. The framework jointly optimizes communication latency, computational load balancing, and hardware utilization while adapting to dynamic workloads. RL learns temporal-action policies for neuron-to-core mapping onto the NoC, and CNN learns the spatial mapping policies to optimize NoC utilization and communication density. By treating DNN layers as computational clusters and learning mapping policies, the framework achieves balanced load distribution and improvements in latency and energy efficiency.