A Scalable Pipelined Time-Domain DTW Engine for Time-Series Classification Using Multibit Time Flip-Flops With 140Giga-Cell-Updates/s Throughput

Zhengyu Chen, Jie Gu

Abstract: Dynamic time warping (DTW), a variant of the dynamic programming algorithm, is widely used for time series classification [1]. Its strong capability for distance measurement for variable-speed temporal sequences makes DTW a popular method for time-series classification in broad applications, such as ECG diagnosis, motion detection, DNA sequencing, etc. [1]. Several efforts have proposed for accelerating the operation of DTW, including a recent demonstration of time-based design in DNA sequencing [2]. However, the demonstration was confined to single-bit operations, a fixed sequence length and low throughput due to nonpipelined operation and a large single-bit delay. To overcome such challenges, this work presents a general-purpose DTW engine for time-series classification using time-domain computing. Pipelined operation is enabled by a time flip-flop (TFF) leading to order-of-magnitude improvements in throughput and a scalable processing capability for time series. Compared with recent time-domain designs, which do not have time-domain memory elements, this work realizes a time-domain pipelined architecture [3].