Playing with number representations and operator-level approximations

International audienceEnergy consumption is one of the major issues in computing today shared by all domains in computer science, from high-performance computing to embedded systems. The two main factors that influence energy consumption is the execution time and data volume. In the recent years, approximation is receiving renewed interests to improve both speed and energy consumption in embedded systems. Many applications in embedded systems do not require high precision/accuracy, and both software designers and hardware designers often seek for a golden point of the compromise between accuracy, speed, energy, and area cost in several layers with a broad range from application, software levels to architecture, circuit levels. Various techniques for approximate computing (AC) augment the design space by providing another set of design knobs for performance-accuracy trade-off. Stochastic computing (SC) is also seen as an alternative to conventional computing, since requiring less hardware and being more tolerant to soft errors at the expense of higher latency. SC uses a probabilistic model of computation and requires less hardware to implement complex operations. This talk will review the main techniques for operator-level approximations using various number representations and by playing with data word-length and types of operators, to show their benefit and drawbacks in terms of energy efficiency. We will also introduce the basic concepts of stochastic computing as well as its advantages in terms of robustness to errors and fair limitations