基于二维脉动阵列的GMM矢量乘法器设计

Design of a GMM Vector Multiplier Based on Two-dimensional Systolic Array

为满足大型识别系统的实时识别请求,需采用基于硬件的GMM说话人识别系统。针对GMM模型的数据通路中吞吐量和精度的瓶颈:矢量乘法器模块和指数运算模块,进行体系结构层次上的研究。本研究针对已有文献中脉动阵列的缺陷,提出一种二维脉动阵列的结构,通过并行处理及插入流水线等方法得到多种体系结构。在速度,面积,吞吐量之间进行折中,应用于各种场合,具有极强的扩展性。最后的测试和分析结果表明,本设计满足标准要求,具有实际的应用价值。

In order to meet the request of large-scale real-time identification recognition system, the hardware-based GMM speaker recognition system is necessary. Aiming at the bottleneck of the throughput and accuracy in the data path for GMM model, vector multiplier module and index operation module, the research on the level of the architecture is conducted. In view of the drawbacks of systolic array existed in existing documents this study proposes a two-dimensional systolic array of the vector multiplier module to obtain multiple architectures with the methods such as parallel processing and inserting pipeline. Through the compromise among speed, area, and throughput, it can be used in various applications and has good expandability. The final verification and synthesis results show that the design fully meet the requirements in the standards, and has actual application value.