Project

DESCRIPTION:

Deep neural networks (DNNs), a prominent framework in machine learning, are characterized by their high computational and energy demands, often involving billions of operations and large working datasets during training. Despite the significant energy and computational resources required for training large DNN models, over time, the costs associated with inference surpass those of training due to the repetitive application of trained models. In fact, it is estimated that inference contributes up to 90% of the energy overhead of a deployed machine learning model. This proportion is anticipated to increase with the growing adoption of AI/ML models in edge computing. Edge-computing devices, which encompass a wide range of hardware including on-board processors in autonomous vehicles, as well as embedded and portable devices used in IoT and wireless applications, stand to benefit greatly from more efficient DNN inference methods. To address this, the proposed project focuses on the development of low-latency, energy-efficient DNN inference accelerators, employing an innovative hardware/software co-design approach. To assess and demonstrate the real-world applicability of this framework, DNN architectures tailored for NextG wireless systems are utilized as the driving application.