Physically-Adaptive Computing via Introspection and Self-Optimization in Reconfigurable Systems.

Abstract

Digital electronic systems typically must compute precise and deterministic results, but in principle have flexibility in how they compute. Despite the potential flexibility, the overriding paradigm for more than 50 years has been based on fixed, non-adaptive inte-grated circuits. This one-size-fits-all approach is rapidly losing effectiveness now that technology is advancing into the nanoscale. Physical variation and uncertainty in com-ponent behavior are emerging as fundamental constraints and leading to increasingly sub-optimal fault rates, power consumption, chip costs, and lifetimes. This dissertation pro-poses methods of physically-adaptive computing (PAC), in which reconfigurable elec-tronic systems sense and learn their own physical parameters and adapt with fine granu-larity in the field, leading to higher reliability and efficiency.

We formulate the PAC problem and provide a conceptual framework built around two major themes: introspection and self-optimization. We investigate how systems can efficiently acquire useful information about their physical state and related parameters, and how systems can feasibly re-implement their designs on-the-fly using the information learned. We study the role not only of self-adaptation—where the above two tasks are performed by an adaptive system itself—but also of assisted adaptation using a remote server or peer.

We introduce low-cost methods for sensing regional variations in a system, including a flexible, ultra-compact sensor that can be embedded in an application and implemented on field-programmable gate arrays (FPGAs). An array of such sensors, with only 1% to-tal overhead, can be employed to gain useful information about circuit delays, voltage noise, and even leakage variations. We present complementary methods of regional self-optimization, such as finding a design alternative that best fits a given system region.

We propose a novel approach to characterizing local, uncorrelated variations. Through in-system emulation of noise, previously hidden variations in transient fault sus-ceptibility are uncovered. Correspondingly, we demonstrate practical methods of self-optimization, such as local re-placement, informed by the introspection data.

Forms of physically-adaptive computing are strongly needed in areas such as com-munications infrastructure, data centers, and space systems. This dissertation contributes practical methods for improving PAC costs and benefits, and promotes a vision of re-sourceful, dependable digital systems at unimaginably-fine physical scales.