Entropy-Aware Self-Optimizing Approximate Multiplier using Operand Distribution Learning
The traditional approximations of multipliers are based on the modes of approximation that are considered to be static or on the accuracy scaling that is controlled externally, and thus, they cannot be used with dynamically changing workloads. This paper suggests an Entropy-Aware Self-Optimizing Approximate Multiplier (ESOM) which will dynamically tune its accuracy, according to real-time learning of the operand distribution. The proposed architecture combines an inconsequential hardware entropy estimator which reads operand sparsity, dynamic range and probability distribution to forecast error tolerance. A micro-architectural learning controller chooses the best approximation modes to reduce the energy used and limit the computational error statistically. In contrast to fixed or manually reconfigurable designs, ESOM automatically responds to input properties, such as workload-sensitive energy-quality tradeoffs. Experimental tests on image processing and neural inference standards prove that there is a major energy saving with regulated degradation in PSNR and SSIM. The proposed design presents a novel paradigm of self-learning arithmetic units of next-generation adaptive edge computing systems.