Approximate Computing in High-Level Synthesis: From Survey to Practical Implementation

Approximate Computing in hardware design has emerged as an alternative way to further reduce the power consumption of integrated circuits (ICs) by trading off errors at the output with simpler, and more efficient logic. So far, the main approaches in approximate computing have been to simplify the hardware circuit by applying different approximation primitives of different aggressiveness to the original hardware description until the maximum error threshold is met. Multiple of these primitives can also be combined to obtain better results. These primitives are often applied at different VLSI design stages to maximize their effect. Because of the importance of this topic, there exists a very large body of work and multiple surveys have tried to cover all of it. In this work, we take a different approach and concentrate only on approximation computing techniques applied at the High-Level Synthesis (HLS) stage of the VLSI design flow. The reason for this is that approximations applied at the highest possible level of VLSI design abstraction also have the highest impact on the resultant circuit. Moreover, HLS is finally being widely embraced by hardware designers, and this work aims at presenting practical examples of how the different approximation primitives can be easily applied using commercial HLS tools. We finally present some typical pitfalls that designers should avoid when using approximate computing and point to some future direction in this area.