Figure 7.

Illustration and process of spike sorting. (A) The training process of the neural spike sorting in PE flow. The neural signals are recorded from analog frontends and stored in SRAM memory. In this instance, the PE computes spike features via PCA and partitions the occurring spikes into clusters via K-means clustering. (B) The inference process of the neural spike sorting in PE and MAC flow. The usage of the MAC unit improves the energy efficiency of spike sorting by around 23.8% compared to performing these spike sorting steps in PE alone.