Data-Driven Approximate Circuit Design

Leverage machine learning techniques to design data-oriented approximate logic circuits.

Electronic devices are becoming smaller and faster each year while embedding more sophisticated functionalities. However, energy efficiency in hardware has become a major bottleneck constraining this trend, especially in the deep nanometer technologies. To this end, approximate computing in hardware design has emerged as one promising approach to ameliorate this issue by slightly sacrificing the accuracy of the computational result in exchange for a large amount of power/area reduction. Another underlying reason behind this computing paradigm is that a large body of prevailing resource-hungry applications such as machine learning and image processing are error-resilient. This indicates that a tolerable amount of error in these applications can be potentially exploited to trade for reductions in hardware complexity.

Approximate computing in hardware design at the logiclevel is broadly divided into two categories according to the design strategies: manual designs and automatic approaches. Manual design strategies have achieved excellent performance on arithmetic elements, however they usually require analyzing the resiliency of each component in a target circuit, which are difficult to generalize to larger circuits. On the other hand, existing automatic design approaches assume a uniform input distribution, which might not always be the case in practice. In particular, many computational tasks are data-oriented such that they mainly operate on a unique data pattern. Therefore, it is imperative to take the data-driven perspective into account when designing approximate logic circuits.

In this project, we aim to close this gap. In specific, we proposed novel machine learning based algorithms to automatically design the compensation circuit, a hardware block used to mitigate computational error, for a given approximate circuit with a specific input distribution. The experimental results proved effectiveness of the proposed approaches, which indeed indicates the need for developing scalable data-driven approximate computing methods.


Towards Data-Driven Approximate Circuit Design
Ling Qiu, Z. Zhang, J. Calhoun, Y. Lao.
Proceedings of IEEE Computer Society Annual Symposium on VLSI (ISVLSI 2019)

A Systematic Method for Approximate Circuit Design Using Feature Selection
Ling Qiu, Y. Lao.
Proceedings of IEEE Computer Society Annual Symposium on VLSI (ISVLSI 2019)