Social Insect-Inspired Adaptive Hardware

Modern VLSI transistor densities allow large systems to be implemented within a single chip. As technologies get smaller, fundamental limits of silicon devices are reached resulting in lower design yields and post-deployment failures. Many-core systems provide a platform for leveraging the computing resource on offer by deep sub-micron technologies and also offer high-level capabilities for mitigating the issues with small feature sizes. However, designing for many-core systems that can adapt to in-field failures and operation variability requires an extremely large multi-objective optimisation space. When a many-core reaches the size supported by the densities of modern technologies (thousands of processing cores), finding design solutions in this problem space becomes extremely difficult. Many biological systems show properties that are adaptive and scalable. This thesis proposes a self-optimising and adaptive, yet scalable, design approach for many-core based on the emergent behaviours of social-insect colonies. In these colonies there are many thousands of individuals with low intelligence who contribute, without any centralised control, to complete a wide range of tasks to build and maintain the colony. The experiments presented translate biological models of social-insect intelligence into simple embedded intelligence circuits. These circuits sense low-level system events and use this manage the parameters of the many-core's Network-on-Chip (NoC) during runtime. Centurion, a 128-node many-core, was created to investigate these models at large scale in hardware. The results show that, by monitoring a small number of signals within each NoC router, task allocation emerges from the social-insect intelligence models that can self-configure to support representative applications. It is demonstrated that emergent task allocation supports fault tolerance with no extra hardware overhead. The response-threshold decision making circuitry uses a negligible amount of hardware resources relative to the size of the many-core and is an ideal technology for implementing embedded intelligence for system runtime management of large-complexity single-chip systems