Approximate Divider Design Based on Counting-Based Stochastic Computing Division
Approximate Divider Design Based on Counting-Based Stochastic Computing Division
复制标题
基于计数随机计算除法的近似除法器设计
DOI：
--
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
S. Tan
中科院分区：
文献类型：
--
作者：
Shuyuan Yu;Yibo Liu;S. Tan
随机计算（SC）在许多新兴应用（如图像处理和深度神经网络）中为容错算术运算提供了极低的成本和能效。然而，现有的基于SC的非线性函数，如除法，需要高度相关的比特流，这与现有的SC计算框架不太适合，在现有的SC计算框架中，需要随机性来获得准确性。在本文中，我们提出了一种新的基于SC的分频器设计的基础上最近提出的基于计数的随机计算方案，这是更准确和更快的比传统的SC，并不依赖于比特流的随机性的准确性。我们展示了如何计数为基础的SC可以应用于非线性函数，如除法。新的分频器，称为计数为基础的分频器，或CBDIV，利用现有的SC为基础的分频方法的相关性要求和高效率的计数为基础的SC计划。它本质上结合了SC中的两个世界中的最好的，并且所得到的除法运算可以作为更有效的部分计数过程来执行。实验结果表明，在32nm工艺节点上实现的CBDIV在精度、延迟、面积、面积延迟积（ADP）和功耗方面分别比现有技术提高了77.8%、37.1%、21.5%、50.6%和25.9%。与定点除法基准相比，CBDIV还节省了31.9%的能耗，并且对于高效图像处理实现所需的二进制输入和输出，CBDIV比现有的基于SC的除法器更节能。此外，具有5位精度的CBDIV甚至可以在精度上超过具有7位精度的最新作品15.4%。最后，我们比较了CBDIV与其他国家的最先进的SC分频器在对比度拉伸应用，并表明CBDIV可以提高精度与20.6dB的平均，这是一个巨大的进步。
Stochastic computing (SC) promises extremely low cost and energy efficiency for error-tolerant arithmetic operations in many emerging applications such as image processing and deep neural networks. Existing SC-based nonlinear functions like division, however, require highly correlated bit-streams, which does not fit well with the existing SC computing framework in which randomness is required for accuracy. In this paper, we propose a novel SC-based divider design based on recently proposed counting-based stochastic computing scheme, which is much more accurate and faster than traditional SC, and does not depend on randomness of bit-streams for accuracy. We show how such counting-based SC can be applied to nonlinear functions like division. The new divider, called counting-based divider, or CBDIV, exploits both the correlation requirement of existing SC-based division methods and high efficiency of counting-based SC scheme. It essentially combines the best of two worlds in SC and the resulting division operation can be performed as a more efficient partial counting process. Experimental results show that the proposed CBDIV implemented in a 32nm technology node outperforms state of art works by 77.8% in accuracy, 37.1% in delay, 21.5% in area, 50.6% in ADP (area delay product) and 25.9% in power. CBDIV also saves 31.9% in energy consumption when compared to the fixed-point division baseline, and is much more energy efficient than existing SC-based dividers for binary inputs and outputs required in efficient image process implementations. Furthermore, CBDIV with 5-bit precision can even outperform state of art works with 7-bit precision in accuracy by 15.4%. Finally, we compare CBDIV with other state of art SC dividers in contrast stretch application and show that CBDIV can improve the accuracy with 20.6dB in average, which is a huge improvement.