Abstract

Approximate computing (AC) has been predominantly recommended for implementation in error-tolerant applications as it offers a reduced resource usage, e.g.,~area and power, for a trade-off in output quality. However, AC implementation has not been adopted in commercial designs yet as it is still falling short in providing a good enough quality. Thus, continued research in the field in the field of improving quality of AC designs is indispensable. In this direction, a recent study exploited the use of machine learning (ML) to improve output quality. Nonetheless, the idea of quality assurance in AC designs could be improved in many aspects.

In the work we present in this thesis, we propose a few practical methods to improve an ML-based quality assurance methodology, which consist of an ML-model that select the most suitable design from a library of AC circuits. For instance, we extend the library of AC designs used for the ML-based approach with larger data path circuits. Larger designs, however, result in an exponential growth of complexity. Thus we propose the use of data pre-processing in order to reduce this hurdle by prioritizing designs based on their physical properties.

Another direction of improving AC circuits designs in general, and the ML-based model in particular is design space exploration (DSE). We therefore propose a novel DSE that drastically reduces the design space based on the aimed targets for area, latency and power of the AC circuit. Moreover, even with a narrowed design space, the number of AC designs to be assessed for their quality could be enormous. Thus, as part of this thesis, we propose a DSE that uses an intricate mathematical modeling for designs to assess their quality.

In another effort in improving quality assurance for AC design, we introduce a highly reliable model that uses a minimal overhead. This work is achieved by using redundant AC modules to form an approximate quadruple modular redundancy (AQMR) design. The proposed AQMR is superior to the exact triple modular redundancy (TMR) by offering a better reliability on top of the resource savings resulting from the implementation of AC.