Energy-Efficient Neural Network Architectures

Abstract

Emerging systems for artificial intelligence (AI) are expected to rely on deep neural networks (DNNs) to achieve high accuracy for a broad variety of applications, including computer vision, robotics, and speech recognition. Due to the rapid growth of network size and depth, however, DNNs typically result in high computational costs and introduce considerable power and performance overheads. Dedicated chip architectures that implement DNNs with high energy efficiency are essential for adding intelligence to interactive edge devices, enabling them to complete increasingly sophisticated tasks by extending battery lie. They are also vital for improving performance in cloud servers that support demanding AI computations.

This dissertation focuses on architectures and circuit technologies for designing energy-efficient neural network accelerators. First, a deep-learning processor is presented for achieving ultra-low power operation. Using a heterogeneous architecture that includes a low-power always-on front-end and a selectively-enabled high-performance back-end, the processor dynamically adjusts computational resources at runtime to support conditional execution in neural networks and meet performance targets with increased energy efficiency. Featuring a reconfigurable datapath and a memory architecture optimized for energy efficiency, the processor supports multilevel dynamic activation of neural network segments, performing object detection tasks with 5.3x lower energy consumption in comparison with a static execution baseline. Fabricated in 40nm CMOS, the processor test-chip dissipates 0.23mW at 5.3 fps. It demonstrates energy scalability up to 28.6 TOPS/W and can be configured to run a variety of workloads, including severely power-constrained ones such as always-on monitoring in mobile applications.

To further improve the energy efficiency of the proposed heterogeneous architecture, a new charge-recovery logic family, called zero-short-circuit current (ZSCC) logic, is proposed to decrease the power consumption of the always-on front-end. By relying on dedicated circuit topologies and a four-phase clocking scheme, ZSCC operates with significantly reduced short-circuit currents, realizing order-of-magnitude power savings at relatively low clock frequencies (in the order of a few MHz). The efficiency and applicability of ZSCC is demonstrated through an ANSI S1.11 1/3 octave filter bank chip for binaural hearing aids with two microphones per ear. Fabricated in a 65nm CMOS process, this charge-recovery chip consumes 13.8µW with a 1.75MHz clock frequency, achieving 9.7x power reduction per input in comparison with a 40nm monophonic single-input chip that represents the published state of the art. The ability of ZSCC to further increase the energy efficiency of the heterogeneous neural network architecture is demonstrated through the design and evaluation of a ZSCC-based front-end. Simulation results show 17x power reduction compared with a conventional static CMOS implementation of the same architecture.