Approximate Computing

Oliver Keszocze
2022 it - Information Technology  
System designers are constantly faced with the task of designing faster systems with a higher throughput, yet, at the same time, lower energy consumption. Artificial Intelligence (AI) applications, most notably in the form of artificial neural networks, are becoming more and more popular and ubiquitous. These applications pose a particular challenge for system designers as they are known for their high demand in computational power resulting in high energy consumption. This greatly hinders the
more » ... dea of Edge Computing, where computations are carried out on cheap, small and often battery powered devices "on the edge", i. e. in the field. For many decades, Moore's law and Dennard scaling led to ever smaller and faster circuits. It seems, that both "laws" have been broken down. Shrinking transistors without negative side-effects is not possible anymore. Parts of modern integrated circuits even need to be turned off during operation in order to meet thermal design power constraints known as "dark silicon". Fortunately, many practical applications can tolerate a certain degree of incorrectness in computations. This can be, for example, due to the limited perception of the human eye, the inherently probabilistic nature of the application or, noisy or redundant input data. Examples for this include image processing where a human observer will not be able notice small deviations in a smoothing operation. Neural networks, for example, are trained to correctly classify input data only to a certain accuracy and, hence, will never produce absolutely accurate results. Designers of GPS devices, know that there is no point in developing a device computing with a higher accuracy than the GPS signal (or the sensors used) can provide.
doi:10.1515/itit-2022-0027 fatcat:2afjt4su45e5bophbdsrotgx6q