Low-power digital CMOS VLSI circuits and design methodologies.

摘要

While improving circuit's speed and reducing its area have been the primary figure of merits in digital VLSI design, more efforts are now spent on minimizing power dissipation. This is becoming equally true for both high-performance chips, such as microprocessors, to reduce cooling costs and improve reliability, as well as portable devices because of their limited energy budget offered by batteries. Over the past few years, the power problem was addressed on all fronts: process, circuits, gates, architectures and systems. This thesis continues this trend by proposing novel low-power techniques and design methodologies at the circuit, gate and architectural levels.; At the circuit level, two new low-swing schemes are presented. The first approach is based on charge sharing and can be used to reduce the swing and so power in dynamic digital circuits with high capacitive loads. Compared to conventional techniques, the proposed approach not only reduces power but also improves the speed as verified by both simulations and measurements. Three application circuits that benefit from this scheme are explored: internal bus lines, match lines in Contents-Addressable Memories (CAMs), and bit-lines in Read-Only Memories (ROMs). For each application, a test chip is fabricated and tested, and measurements have confirmed the functionality and high speed down to the low-voltage region of operation. The second low-swing circuit technique is based on current-injection. This approach is applied to the write and read operations in multi-port SRAM cell design. Simulations have shown the superiority of the approach in terms of both speed and power as compared to conventional ones. A test chip of a 3-port register file is fabricated and testing results have proved the applicability of the technique.; At the gate-level, we propose DVDV; a new automated design approach for reducing power dissipation in high-speed deep sub-micron CMOS Logic circuits. The main idea is to utilize a library of gates having Dual supply voltages (V_dd) and Dual threshold voltages (V _th), hence the name DVDV, to achieve high-speed at the lowest possible dynamic and leakage power dissipation. A simple algorithm for DVDV technology mapping is developed and implemented in C under the Berkeley's SIS-1.2 environment. Application to benchmark circuits shows large power savings and speed improvement as compared to using two supply voltages (and a single V_th) or two threshold voltages (and a single V_dd).; At the architectural-level, a method to characterize the effective capacitance in data path macros is developed. Given a library of hard-macros, a capacitance model based on multiple linear regression is derived for each macro. The capacitance models can be used later during architectural-level power estimation. The characterization methodology assumes no specific input data statistics, requires little knowledge about the module structure, allows the user to trade-off accuracy and characterization time, and propagates power from transistor-level (real) implementations. Simulation experiments on a set of data-path components show accuracy within 15% from a transistor-level tool on the average.