Generic architectures for efficient Hyper-Dimensional Computing
Skip to main content
eScholarship
Open Access Publications from the University of California

UC Berkeley

UC Berkeley Electronic Theses and Dissertations bannerUC Berkeley

Generic architectures for efficient Hyper-Dimensional Computing

Abstract

The last decade has witnessed a slowdown in technology scaling. At the same time, the emergence of machine learning has substantially increased computational demand. While these trends have seriously challenged traditional paradigms for digital design, novel computing methods based on randomness can be leveraged for continued increase in performance and energy efficiency. Hyper-dimensional Computing (HDC), a brain-inspired paradigm using high-dimensional random vectors as its fundamental data-type, shows promise. It is known to provide competitive accuracy on sequential prediction tasks with far smaller model size and training time compared to conventional machine learning, and is robust to representation errors. This dissertation considers efficient architectures for real-time Hyper-dimensional Computing on edge devices. After a study of implementations on classical compute platforms, a highly-pipelined, data-flow architecture is developed from the first principles of HDC. A detailed construction for a reasonably large HDC processor, capable of supporting a wide variety of prominent body-sensing applications, forms the foundation of this work. The basic architecture is extended to HDC processors requiring more than a single bit of representation for vector elements. Using results from high-dimensional probability theory, a numerical normalization is proposed and its effectiveness is proven for applications obeying reasonable assumptions on vector elements’ distribution. Verification experiments indicate that empirical performance of the proposed normalization is far better than the theoretical guarantee. A 2048-bit wide HDC processor, designed using the architectural principles developed here, was manufactured in a leading technology node. The chip reliably achieves an energy cost of approximately 30 nanojoules per classification on a state-of-the-art body-sensing application – the best when compared to all known previous hardware in the existing literature. Measurements establish energy efficiency and robustness of the designed processor. The architecture, arguments, experiments and measurements presented in this dissertation confirm the great potential of Hyper-Dimensional Computing as a computing paradigm capable of competitive performance in extremely energy-constrained environments.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View