A programmable analog calculation unit for vector computations

Abstract

In this paper, a programmable analog calculation unit is designed for computing arbitrary functions of highly dimensional vectors in real-time. The support vector regression algorithm with Gaussian kernels is employed to approximately retrieve vector calculations. Through the proposed efficient scheme of regression, the number of support vectors is greatly reduced. A compact analog circuit is implemented to generate dynamically tunable Gaussian kernel functions of vectors. By integrating the kernel generation circuits in fully parallel, arbitrary vector functions are retrieved without any clock-cycle latency. The proof-of-concept processor is designed in a standard 0.18 um CMOS technology for vector calculations with the maximum dimensions of nine, which is implemented by only 9460 transistors. From the circuit simulation results, the proposed calculation unit successfully retrieves all the example functions in real-time with the maximum inaccuracy of 2.92%, 4.2%, and 10.85% for one, two, and nine dimensional regressions, respectively.