An efficient power and accuracy trade-off in error-resilient applications using approximate multipliers with error correction

High-performance computing depends on immense computational power to address complex problems, yet this often entails significant energy and resource costs. Approximate computing provides a complementary solution by deliberately relaxing precision in various calculations, such as neural network computations, thereby enabling faster execution and reduced energy consumption. In this regard, the multiplier is a critical component in the arithmetic units of convolutional neural networks (CNNs) and image processing systems. Given the high volume of multiplication and accumulation operations in CNNs, employing low-power, area-efficient approximate multipliers can significantly enhance overall performance. In this paper, we evaluate two recently proposed 8-bit approximate multipliers, each incorporating an error correction block, and compare them with their exact and approximate counterparts. These multipliers are assessed based on error parameters both with and without error correction. As an application in image processing, the proposed designs are implemented in the multiplication operations for image sharpening, achieving a mean structural similarity index of 98%. Furthermore, we employ the 8-bit approximate multiplier in several neural networks to evaluate classification accuracy. The accuracy levels achieved by LeNet-5, VGG-16, and ResNet-18 are above 98, 82, and 86%, respectively. We consider a trade-off between performance parameters and classification accuracy for datasets. The results of the figure-of-merit analysis demonstrate that using those two multipliers leads to an improvement of up to 89% compared with the exact design.