Advancing AI Hardware: Asynchronous Tsetlin Machines for Enhanced Energy Efficiency

Discover how the asynchronous Tsetlin Machine design revolutionises energy efficiency in AI, making pervasive artificial intelligence a reality.

Authors: include Adrian Wheeldon, Alex Yakovlev, Rishad Shafik

Published: 2021, IEEE International Symposium on Asynchronous Circuitsty

Summary

Advancing the frontier of pervasive artificial intelligence, a novel asynchronous design for Tsetlin Machines presents a breakthrough in low-energy, high-efficiency AI hardware.

This paper introduces a pioneering hardware design tailored for the Tsetlin Machine, focusing on asynchronous learning and inference pathways to enhance energy efficiency and reduce latency. The innovative use of asynchronous design techniques, such as Petri nets, signal transition graphs, dual-rail, and bundled-data systems, underpins the creation of low-energy hardware capable of supporting pervasive AI applications. This architecture significantly improves the practicality of AI systems in energy-constrained environments such as personalized healthcare devices and IoT ecosystems.

The core achievement of this design is its detailed breakdown of the automaton feedback, probability generation, and Tsetlin automata, culminating in a latency analysis of the inference datapath. The research showcases the practical advantages of asynchronous design, particularly in applications where latency and energy efficiency are critical. Moreover, the paper addresses the technical challenges involved in static timing analysis within asynchronous circuits, providing insights into overcoming these hurdles in system implementation.

By blending cutting-edge hardware design with the inherent efficiency of Tsetlin Machines, this research paves the way for the deployment of intelligent systems where power availability is limited and operational demands are high, setting a new benchmark for the integration of AI into daily technological applications.

Read more

The full paper Self-timed reinforcement learning using Tsetlin Machine is available from IEEE International Symposium on Asynchronous Circuitsty.

Article

First published by IEEE International Symposium on Asynchronous Circuitsty on 27 January 2021.

DOI: arXiv:2109.00846