Algorithm/Architecture Co-Design for Low-Power Neuromorphic Computing

Abstract

The development of computing systems based on the conventional von Neumann architecture has slowed down in the past decade as complementary metal-oxide-semiconductor (CMOS) technology scaling becomes more and more difficult. To satisfy the ever-increasing demands in computing power, neuromorphic computing has emerged as an attractive alternative. This dissertation focuses on developing learning algorithm, hardware architecture, circuit components, and design methodologies for low-power neuromorphic computing that can be employed in various energy-constrained applications.

A top-down approach is adopted in this research. Starting from the algorithm-architecture co-design, a hardware-friendly learning algorithm is developed for spiking neural networks (SNNs). The possibility of estimating gradients from spike timings is explored. The learning algorithm is developed for the ease of hardware implementation, as well as the compatibility with many well-established learning techniques developed for classic artificial neural networks (ANNs). An SNN hardware equipped with the proposed on-chip learning algorithm is implemented in CMOS technology. In this design, two unique features of SNNs, the event-driven computation and the inferring with a progressive precision, are leveraged to reduce the energy consumption. In addition to low-power SNN hardware, accelerators for ANNs are also presented to accelerate the adaptive dynamic programing algorithm. An efficient and flexible single-instruction-multiple-data architecture is proposed to exploit the inherent data-level parallelism in the inference and learning of ANNs. In addition, the accelerator is augmented with a virtual update technique, which helps improve the throughput and energy efficiency remarkably. Lastly, two techniques in the architecture-circuit level are introduced to mitigate the degraded reliability of the memory system in a neuromorphic hardware owing to the aggressively-scaled supply voltage and integration density. The first method uses on-chip feedback to compensate for the process variation and the second technique improves the throughput and energy efficiency of a conventional error-correction method.