FM-AM: Fused-Metrics Approximate Multiplier Exploration Framework for DNNs

Quantization combined with approximate computing has been extensively explored in Deep Neural Networks (DNNs) for edge devices and hardware accelerators, offering significant reductions in memory footprint and computation overhead. However, effectively integrating these techniques presents two key challenges: (1) Conventional error metrics, such as mean squared error, fail to accurately capture the proper tradeoff between network accuracy and hardware efficiency. (2) The design space for mixed-precision DNNs is vast, making exhaustive exploration computationally prohibitive. To address these challenges, we propose the Fused-Metrics Approximate Multiplier Exploration (FM-AM) framework, which introduces two key innovations: (1) A cross-layer optimization approach that jointly configures quantization bit-widths and energy-efficient approximate multipliers, guided by fused-metrics error estimation and hardware constraints. (2) A Bayesian Optimization-based search algorithm that efficiently selects approximate computing units by pruning the multiplier library, significantly reducing search time while meeting accuracy requirements. We evaluate FM-AM on AlexNet, ResNet-18, and ResNet-50 using the CIFAR-100 dataset. Experimental results demonstrate energy reductions of 60.2%, 58.9%, and 50.0%, respectively, with only 1.63%, 1.59%, and 1.34% accuracy degradation when implemented in a 28-nm CMOS industrial technology.