The CNNUC3: an analog I/O 64x64 CNN universal machine chip prototype with 7-bit analog accuracy
Proceedings of the 2000 6th IEEE International Workshop on Cellular Neural Networks and their Applications (CNNA 2000) (Cat. No.00TH8509)
This paper describes a full-custom mixed-signal chip which embeds distributed optical signal acquisition, digitally-programmable analog parallel processing, and distributed image memorycache -on a common silicon substrate. This chip, designed in a 0Spm CMOS standard technology contains around 1. OOO. OOO transistors, 80% of which operate in analog mode; it is hence one the most complex mixed-signal chip reported to now. Chip functional features are in accordance to the CNN Universal Machine [l]
... iversal Machine [l] paradigm cellular, spatial-invariant array architecture; programmable local interactions among cells; randomly-selectable memory of instructions (elementary instructions are defined by specific values of the cell local interactions); random storagdretrieval of intermediate images; capability to complete algorithmic image processing tasks controlled by the user-selected stored instructions and interacting with the cache memory, etc. Thus, as illustrated in this paper, the chip is capable to complete complex spatio-temporal image processing tasks within short computation time ( -ZWns for linear convolutions) and using a low power budget (42W for the complete chip). The internal circuitq of the chip has been designed to operate in robust manner with >7-bit equivalent accuracy in the internal analog operations, which has been confirmed by experimental measurements. Hence, to all practical purposes, processing tasks completed by the chip have the same accuracy than those completed by digital processors preceded by 7-bit digital-to-analog converters for image digitalization. Such 7-bit accuracy is enough for most image processing applications. CNNUC3 has been demonstrated capable to implementeither directly or through template decomposition -100% of the linear 3 x 3 templates in reported . General Characteristics CNNUC3 consists basically of an m a y of 64 x 64 identical cells. Its processing is continuous-time and spa-Feedback and control templates, and the offset (or bias) term are programmable with a resolution of eight bits tially-invariant, with radius-1 neighbourhood and the cell state equation given by the FSR model . 7. This work hm been partially funded by ONR-NICOP N68171-98-C-9004, DICTAM IST-19w-lw07 and TIC 990826.