Exploring Asynchronous Design Techniques for Process-Tolerant and Energy-Efficient Subthreshold Operation

Abstract

Supply voltage scaling is one of the easiest ways to reduce energy dissipation. Therefore, researchers have considered subthreshold logic as a promising option to achieve ultra low energy dissipation. However, circuit propagation delay is extremely sensitive to PVT variations under subthreshold operation. Hence, large delay margin is required for successful operation of conventional synchronous designs. Since leakage energy contributes to a substantial portion of total energy dissipation in subthreshold operation, the leakage energy dissipated for the required delay margin degrades energy efficiency significantly. In addition, even small intra-die variations result in large clock skew and hence, it is difficult to efficiently handle timing issues such as the setup and the hold time violations. In this work, we explore asynchronous design approach to address these challenges in subthreshold operation. We employ critical-path replica to generate completion signals of combinational logic blocks and use classical four-phase handshaking for communication between pipeline flip-flops. Since the proposed design approach uses only local clock buffers, it is easier to handle timing problems compared to synchronous designs. We compared iso-yield minimum energy dissipation of two design approaches (synchronous and asynchronous) in an inverter chain. Despite leakage overhead due to pad delay of critical-path delay line and ?return-to-zero? time of four-phase handshaking, the proposed asynchronous design shows 71% energy savings compared to its synchronous counterpart. To demonstrate subthreshold operation of the proposed design approach, we fabricated an 8-tap FIR filter in 90 nm CMOS. Measured oscilloscope plots of handshaking and output bus signals show that the design operates successfully below 300 mV. We also measured energy consumption of the FIR filter from 19 test chips-the average was 4.64 pJ and the standard deviation was 0.3526 pJ.