AMORL: A Training-Aware, Adaptive, Multi-Objective RL-CNN Framework for Dynamic DNN Mapping on Networks-on-Chip
Year: 2026
Type: preprint
Abstract: 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, par-allelism, and power efficiency over traditional buses. These properties enable efficient execution of computations and communications for deep neural network (DNN) layers on NoCs. However, traditional mapping strategies struggle with the diverse communication patterns of DNNs, such as inter-layer communication and one-to-many/many-to-one patterns in fully connected and convolutional layers, necessitating mapping techniques that mitigate NoC communication bottlenecks. This work proposes AMORL, a generalized reinforcement-learning and convolutional-neural-network (RL–CNN) framework for scalable, adaptive mapping of DNN workloads onto heterogeneous NoC architec-tures. RL learns temporal neuron-to-core mapping policies, while CNN captures spatial communication structure to guide locality-aware decisions. The framework jointly optimizes communication latency, load balancing, and hardware utilization under dynamic workloads. Evaluations across six DNN workloads, ResNet-50, BERT-tiny, MNIST, CIFAR-10, IRIS, and Face Recognition, on NoC sizes from 4 × 4 to 32 × 32 demonstrate that AMORL reduces communication cost by over 70%, load imbalance by up to 86%, and latency by more than 50% compared to ablated variants, while maintaining energy-efficient operation. Cycle-accurate gem5/Garnet simulations on 4 × 4 and 8 × 8 meshes further validate AMORL’s mappings, showing sustained throughput up to 876 Gbps on the 4 × 4 NoC and over 2.77 Tbps on the 8 × 8 NoC, confirming the framework’s robustness under realistic traffic conditions.
Source: Research Square
Author: Md Farhadur Reza
Institution: Eastern Illinois University
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204158831