Energy-efficient Mixed-mode support vector machine processor with analog Gaussian kernel

A Mixed-mode SVM processor is proposed for energy efficient SVM classification. Analog Gaussian kernel datapath enables the energy efficient and reconfigurable SVM processing with a digital controller. Digitally-assisted calibration shapes the Gaussian curve with the pre-defined width to remove non-uniformity under temperature variation and enhance the classification accuracy. Completion detection of the analog Gaussian circuit optimally controls the 2 stage pipeline with the analog datapath and the digital controller. As a result, the implemented chip achieves the processing speed of 17.5Mvectors/s, enhanced by 47% and the energy efficiency of 22.6pJ/vectors*input dimension.