Circuit Techniques for Adaptive and Reliable High Performance Computing.

Abstract

Increasing power density with process scaling has caused stagnation in the clock speed of modern microprocessors. Accordingly, designers have adopted message passing and shared memory based multicore architectures in order to keep up with the rapidly rising demand for computing throughput. At the same time, applications are not entirely parallel and improving single-thread performance continues to remain critical. Additionally, reliability is also worsening with process scaling, and margining for failures due to process and environmental variations in modern technologies consumes an increasingly large portion of the power/performance envelope. In the wake of multicore computing, reliability of signal synchronization between the cores is also becoming increasingly critical. This forces designers to search for alternate efficient methods to improve compute performance while addressing reliability. Accordingly, this dissertation presents innovative circuit and architectural techniques for variation-tolerance, performance and reliability targeted at datapath logic, signal synchronization and memories.

Firstly, a domino logic based design style for datapath logic is presented that uses Adaptive Robustness Tuning (ART) in addition to timing speculation to provide up to 71% performance gains over conventional domino logic in 32bx32b multiplier in 65nm CMOS. Margins are reduced until functionality errors are detected, that are used to guide the tuning.

Secondly, for signal synchronization across clock domains, a new class of dynamic logic based synchronizers with single-cycle synchronization latency is presented, where pulses, rather than stable intermediate voltages cause metastability. Such pulses are amplified using skewed inverters to improve mean time between failures by ~1e6x over jamb latches and double flip-flops at 2GHz in 65nm CMOS.

Thirdly, a reconfigurable sensing scheme for 6T SRAMs is presented that employs auto-zero calibration and pre-amplification to improve sensing reliability (by up to 1.2 standard deviations of NMOS threshold voltage in 28nm CMOS); this increased reliability is in turn traded for ~42% sensing speedup.

Finally, a main memory architecture design methodology to address reliability and power in the context of Exascale computing systems is presented. Based on 3D-stacked DRAMs, the methodology co-optimizes DRAM access energy, refresh power and the increased cost of error resilience, to meet stringent power and reliability constraints.