*6.1.1 UCR time-series clustering benchmark suite*

Using *TNNSim*, we designed 36 single-(mini) column TNNs of various sizes to perform unsupervised clustering on 36 UCR time-series datasets [58]. These singlecolumn TNNs can achieve state-of-the-art performance surpassing most of the contemporary comparable algorithms [13]. **Table 1** provides synapse count and PPA for 7 of the 36 benchmark-specific designs to demonstrate the efficacy of single minicolumn designs across various configurations and various benchmark applications. The seven benchmarks and their input signal types are briefly described: (1) *SonyAIBORobotSurface2* - accelerometer signals to detect two types of walking surfaces; (2) *ECG200* - ECG signals to detect normal heartbeat vs. myocardial infarction; (3) *Wafer* - fabrication process control sensor signals to detect normal vs. abnormal silicon wafers; (4) *ToeSegmentation2* - motion sensors to detect normal vs. abnormal walking; (5) *Lightning2* - power-density series derived from optical and RF sensor spectogram to detect lightning; (6) *Beef* - food spectograph to detect varying levels of adulteration; and (7) *WordSynonyms* - 1D series from word outlines to detect 25 different words.

The smallest mini-column (130 synapses) for *SonyAIBORobotSurface2* and the largest mini-column (6750 synapses) for *WordSynonyms* perform very efficient unsupervised clustering within only 1 *μ*W and 40 *μ*W power, respectively [12]. PPA metrics scale with synapse count as expected. Note that mini-column for *Beef* has the highest input synapse count *p* and therefore incurs the largest computation time, as delay depends on *p*. These benchmarks demonstrate even relatively small single minicolumn designs can perform practically useful clustering with minimal hardware complexity and power consumption.
