Title: Methodology and Analysis for Efficient Custom Architecture Design using Machine Learning
Dr. Tushar Krishna, ECE, Chair, Advisor
Dr. , Co-Advisor
Dr. Saibal Mukhopadhyay, ECE
Dr. Hyesoon Kim, CoC
Dr. Vivek Sarkar, CoC
Dr. Vijay Reddi, Harvard
Abstract: Machine learning algorithms especially Deep Neural Networks (DNNs) have revolutionized the arena of computing in the last decade. DNNs along with the computational advancements also bring in an unprecedented appetite for compute and parallel processing. Computer architects have risen to challenge by creating novel custom architectures called accelerators. However, given the ongoing rapid advancements in algorithmic development accelerators architects are playing catch-up to churn out optimized designs each time new algorithmic changes are published. It is also worth noting that the accelerator design cycle is expensive. It requires multiple iterations of design space optimization and expert knowledge of both digital design as well as domain knowledge of the workload itself. It is therefore imperative to build scalable and flexible architectures which are adaptive to work well for a variety of workloads. Moreover, it is also important to develop relevant tools and design methodologies that lower the overheads incurred at design time such that subsequent design iterations are fast and sustainable. This thesis takes a three-pronged approach to address these problems and push the frontiers for the DNN accelerator design process. First, the thesis presents the description of a now popular cycle-accurate DNN accelerator simulator. This simulator is built to obtain detailed metrics as fast as possible. A detailed analytical model is also presented in this thesis which enables the designer to understand the interactions of the workload and architecture parameters. The information from the model can be directly used to prune the design search space to achieve faster convergence. Second, the thesis details a couple of flexible yet scalable DNN accelerator architectures. Finally, this thesis describes the use of machine learning to capture the design space of DNN accelerators and train a model to predict optimum configurations when queried with workload parameters and design constraints. The novelty of this piece of work is that it systematically lays out the formulation of traditional design optimization into a machine learning problem and also describes the quality and components of a model which works well across various architecture design tasks.