Study and development of innovative strategies for energy-efficient cross-layer design of digital VLSI systems based on Approximate Computing

The increasing demand on requirements for high performance and energy efficiency in modern digital systems has led to the research of new design approaches that are able to go beyond the established energy-performance tradeoff. Looking at scientific literature, the Approximate Computing paradigm has been particularly prolific. Many applications in the domain of signal processing, multimedia, computer vision, machine learning are known to be particularly resilient to errors occurring on their input data and during computation, producing outputs that, although degraded, are still largely acceptable from the point of view of quality. The Approximate Computing design paradigm leverages the characteristics of this group of applications to develop circuits, architectures, algorithms that, by relaxing design constraints, perform their computations in an approximate or inexact manner reducing energy consumption. This PhD research aims to explore the design of hardware/software architectures based on Approximate Computing techniques, filling the gap in literature regarding effective applicability and deriving a systematic methodology to characterize its benefits and tradeoffs. The main contributions of this work are: -the introduction of approximate memory management inside the Linux OS, allowing dynamic allocation and de-allocation of approximate memory at user level, as for normal exact memory; - the development of an emulation environment for platforms with approximate memory units, where faults are injected during the simulation based on models that reproduce the effects on memory cells of circuital and architectural techniques for approximate memories; -the implementation and analysis of the impact of approximate memory hardware on real applications: the H.264 video encoder, internally modified to allocate selected data buffers in approximate memory, and signal processing applications (digital filter) using approximate memory for input/output buffers and tap registers; -the development of a fully reconfigurable and combinatorial floating point unit, which can work with reduced precision formats