Abstract

Shifting the compute workloads from cloud toward edge devices can significantly improve the overall latency for inference and learning. On the contrary this paradigm shift exacerbates the resource constraints on the edge devices. Neuromorphic computing architectures, inspired by the neural processes, are natural substrates for edge devices. They offer co-located memory, in-situ training, energy efficiency, high memory density, and compute capacity in a small form factor. Owing to these features, in the recent past, there has been a rapid proliferation of hybrid CMOS/Memristor neuromorphic computing systems. However, most of these systems offer limited plasticity, target either spatial or temporal input streams, and are not demonstrated on large scale heterogeneous tasks. There is a critical knowledge gap in designing scalable neuromorphic systems that can support hybrid plasticity for spatio-temporal input streams on edge devices.

This research proposes Pyragrid, a low latency and energy efficient neuromorphic computing system for processing spatio-temporal information natively on the edge. Pyragrid is a full-scale custom hybrid CMOS/Memristor architecture with analog computational modules and an underlying digital communication scheme. Pyragrid is designed for hierarchical temporal memory, a biomimetic sequence memory algorithm inspired by the neocortex. It features a novel synthetic synapses representation that enables dynamic synaptic pathways with reduced memory usage and interconnects. The dynamic growth in the synaptic pathways is emulated in the memristor device physical behavior, while the synaptic modulation is enabled through a custom training scheme optimized for area and power.

Pyragrid features data reuse, in-memory computing, and event-driven sparse local computing to reduce data movement by ~44x and maximize system throughput and power efficiency by ~3x and ~161x over custom CMOS digital design. The innate sparsity in Pyragrid results in overall robustness to noise and device failure, particularly when processing visual input and predicting time series sequences. Porting the proposed system on edge devices can enhance their computational capability, response time, and battery life.

Library of Congress Subject Headings

Machine learning; Neural networks (Computer science); Computer architecture; Memristors; Metal oxide semiconductors, Complementary

Publication Date

9-2020

Document Type

Dissertation

Student Type

Graduate

Degree Name

Engineering (Ph.D.)

Department, Program, or Center

Computer Engineering (KGCOE)

Advisor

Dhireesha Kudithipudi

Advisor/Committee Member

Santosh Kurinec

Advisor/Committee Member

Cory Merkel

Campus

RIT – Main Campus

Plan Codes

ENGR-PHD

Share

COinS