Rapid Emulation of Approximate DNN Accelerators

In recent years, Deep Neural Networks (DNNs) have become essential tools, surpassing human capabilities in various applications. To address the computational complexity of DNNs, specialized hardware accelerators and Approximate Computing (AC) techniques have emerged. This paper presents a novel stochastic method for efficiently emulating Approximate Multipliers (AMs) within DNNs. This method offers a significant 4× speedup compared to the fastest GPU-based platform and a remarkable 28.25× speedup compared to the latest CPU-based platform with variable bit widths. It serves as a valuable tool for selecting viable AM candidates and provides valuable insights, particularly in scenarios involving a large number of AMs.