Orthogonal tRAM and bank-driven embedded memory architecture for two dimensional signal processing

Abstract

The embedded memories in the embedded system are prevalent with advances in the VLSI technology and the associated design paradigms have shifted. In this thesis, embedded memory architectures for two-dimensional signal processing have been proposed for efficient data flow in the I/O bounded problems and optimized under the off-/on-chip memory bandwidth and one-chip criterion. The figure of merits will be addressed through the intrinsic embedded memory architecture and the usage in the application such as Motion Estimation (ME). Unlike the off-chip memory increasing the peak performance, embedded memories can be exploited to suit specific applications for the two-dimensional signal processing where overhead clocks are induced by the architectural hazards in mismatch between storage components and Processing Elements (PEs). By applying an architecture driven voltage scale, we analyze the effects of the overhead clocks in power consumptions. For reducing the power consumed by these overhead clocks and maximizing the useful data transfer rate, we will propose two approaches using the asymmetric/complementary embedded memory banks and the orthogonal transpose-RAM cell array. The prototype processor with asymmetric, complementary embedded memory banks has been fabricated with 0.6um standard CMOS technology with 1-poly and 3-metal, where several techniques are employed to reduce the silicon area occupied by the memory block and room for the metal tracks used in PEs. The results show that power saving is improved by using complementary access types of memory banks and amounts to 27.3% compared to an identical design without the proposed enhancements when the Full-search Block Matching Algorithm (FBMA) is applied for the CCIR-601 format. The proposed asymmetric, complementary embedded memory banks make an ME processor efficient and optimal for interest wide range in spite of the area-overhead. The implemented chip operates at 2.5mW/MHz and 3.3V supply voltage. Accordi...