A 384-Channel Online-Spike-Sorting IC Using Unsupervised Geo-OSort Clustering and Achieving 0.0013mm<sup>2</sup>/Ch and $1.78\mu \text{W/ch}$
Yingping Chen, Bernardo Tacca, Yunzhu Chen, Dwaipayan Biswas, Georges Gielen, Francky Catthoor, Marian Verhelst, Carolina Mora López
Abstract
As neural-recording devices get denser and generate more data [1], on-chip and online neural-signal processing becomes crucial to reduce the data-transmission power and enable real-time closed-loop applications with minimum latency. Spike sorting (SS) is an important data-reduction technique that allows the classification of extracellularly recorded spikes into clusters representing the neuron sources. SS requires computationally intensive algorithms usually implemented offline in software [2]. However, to enable on-chip and online SS (OSS), a low-complexity and hardware-efficient algorithm is required to achieve minimal area, energy and latency, while maintaining comparable accuracy. Although several multi-channel OSS chips have been reported recently [3]–[7], they either require offline [3], [6] or on-chip training [4], [5], [7], consume high power [5] or area [5], [6], or achieve low accuracy [4], [6] or high latency [4], [6]. Moreover, they struggle to balance the trade-off between computational complexity and performance, limiting the number of channels to <128. To overcome these limitations, this work proposes 3 key techniques for area-and energy-efficient hardware OSS, achieving software-comparable accuracy with 384 channels.