Algorithm And Architecture Design Of Network On Chip Based Convolutional Neural Network Accelerator

Author : Mahmood Azhar Qureshi
Publisher :
ISBN 13 :
Total Pages : 0 pages
Book Rating : 4.:/5 (133 download)

DOWNLOAD NOW!

Download or read book Design of High-performance and Energy-efficient Accelerators for Convolutional Neural Networks written by Mahmood Azhar Qureshi and published by . This book was released on 2021 with total page 0 pages. Available in PDF, EPUB and Kindle. Book excerpt: Deep neural networks (DNNs) have gained significant traction in artificial intelligence (AI) applications over the past decade owing to a drastic increase in their accuracy. This huge leap in accuracy, however, translates into a sizable model and high computational requirements, something which resource-limited mobile platforms struggle against. Embedding AI inference into various real-world applications requires the design of high-performance, area, and energy-efficient accelerator architectures. In this work, we address the problem of the inference accelerator design for dense and sparse convolutional neural networks (CNNs), a type of DNN which forms the backbone of modern vision-based AI systems. We first introduce a fully dense accelerator architecture referred to as the NeuroMAX accelerator. Most traditional dense CNN accelerators rely on single-core, linear processing elements (PEs), in conjunction with 1D dataflows, for accelerating the convolution operations in a CNN. This limits the maximum achievable ratio of peak throughput per PE count to unity. Most of the past works optimize their dataflows to attain close to 100% hardware utilization to reach this ratio. In the NeuroMAX accelerator, we design a high-throughput, multi-threaded, log-based PE core. The designed core provides a 200% increase in peak throughput per PE count while only incurring a 6% increase in the hardware area overhead compared to a single, linear multiplier PE core with the same output bit precision. NeuroMAX accelerator also uses a 2D weight broadcast dataflow which exploits the multi-threaded nature of the PE cores to achieve a high hardware utilization per layer for various dense CNN models. Sparse convolutional neural network models reduce the massive compute and memory bandwidth requirements inherently present in dense CNNs without a significant loss in accuracy. Designing sparse accelerators for the processing of sparse CNN models, however, is much more challenging compared to the design of dense CNN accelerators. The micro-architecture design, the design of sparse PEs, addressing the load-balancing issues, and the system-level architectural design issues for processing the entire sparse CNN model are some of the key technical challenges that need to be addressed in order to design a high-performance and energy-efficient sparse CNN accelerator architecture. We break this problem down into two parts. In the first part, using some of the concepts from the dense NeuroMAX accelerator, we introduce SparsePE, a multi-threaded, and flexible PE, capable of handling both the dense and sparse CNN model computations. The SparsePE core uses the binary mask representation to actively skip ineffective sparse computations involving zeros, and favors valid, non-zero computations, thereby, drastically increasing the effective throughput and the hardware utilization of the core as compared to a dense PE core. In the second part, we generate a two-dimensional (2D) mesh architecture of the SparsePE cores, which we refer to as the Phantom accelerator. We also propose a novel dataflow that supports processing of all layers of a CNN, including unit and non-unit stride convolutions (CONV), and fully-connected (FC) layers. In addition, the Phantom accelerator uses a two-level load balancing strategy to minimize the computational idling, thereby, further improving the hardware utilization, throughput, as well as the energy efficiency of the accelerator. The performance of the dense and the sparse accelerators is evaluated using a custom-built cycle accurate performance simulator and performance is compared against recent works. Logic utilization on hardware is also compared against the prior works. Finally, we conclude by mentioning some more techniques for accelerating CNNs and presenting some other avenues where the proposed work can be applied.