Challenges in mixed-signal IC design of CNN chips in submicron CMOS

Abstract

Summary form only given. The contrast observed between the performance of artificial vision machines and "natural" vision system is due to the inherent parallelism of the former. In particular, the retina combines image sensing and parallel processing to reduce the amount of data transmitted for subsequent processing by the following stages of the human vision system. Industrial applications demand CMOS vision chips capable of flexible operation, with programmable features and standard interfacing to conventional equipment. The CNN Universal Machine (CNN-UM) is a powerful methodological framework for the systematic development of these chips. Basic system-level targets in the design of these chips are to increase the cell density and operation speed. As the technology scales down to submicron all the lateral dimensions decrease by the scaling factor /spl lambda/, and the vertical dimensions scale as /spl lambda//sup -a/, where a is typically around 1/2. Ideally, cell density /spl prop//spl lambda//sup 2/ and time constant /spl prop//spl lambda//sup -2/. The article explains why this is not strictly true, and addresses the challenges involved in the design of CNN chips in submicron technologies.