TEMET: Truncated REconfigurable Multiplier with Error Tuning

Approximate computing is a well-established technique to mitigate power consumption in error-tolerant domains such as image processing and machine learning. When paired with reconfigurable hardware, it enables dynamic adaptability to each specific task with improved power-accuracy trade-offs. In this work, we present a design methodology to enhance the energy and error metrics of a signed multiplier. This novel approach reduces the approximation error by leveraging a statistic-based truncation strategy. Our multiplier features 256 dynamically configurable approximation levels and run-time selection of the result precision. Our technique improves the mean-relative error by up to 34% compared to the zero truncation mechanism. Compared with an exact design, we achieve a maximum of 60.1% power saving for a PSNR of 10.3dB on a 5x5 Sobel filter. Moreover, we reduce the computation energy of LeNet by 31.5%, retaining 89.4% of the original accuracy on FashionMNIST.