A novel low overhead fault tolerant Kogge-Stone adder using adaptive clocking
Proceedings of the conference on Design, automation and test in Europe
Trifecta: a nonspeculative scheme to exploit common, data-dependent subcritical paths
IEEE Transactions on Very Large Scale Integration (VLSI) Systems
| Hi-index | 0.00 |
Bridging defects typically manifest themselves as increased path delays instead of stuck-at failures. On the other hand, parametric variations (both inter- and intra-die) increase the spread of the circuit delay. Low power design techniques such as voltage scaling, dual-Vth etc. deteriorate the delay spread further. These mechanisms for delay variations in nanoscaled technologies significantly affect the parametric yield. We propose a new design methodology to tolerate subtle delay failures that arise both due to manufacturing defects and parameter fluctuations. We synthesize the circuit to (a) isolate and predict the critical paths of a circuit; (b) create timing slack between critical and off-critical paths and ensure that they are activated rarely; and, (c) avoid the delay failures in these paths by adaptively stretching the clock period. Since critical paths are the most sensitive section of the circuit in terms of delay defects, we ensure fault-free operation by isolating them and providing extra computation time by predicting their activation. This allows us to achieve the required yield with small performance penalty (due to occasional clock stretching under critical path activation). We present application of the proposed methodology for both linear and non-linear pipeline designs. We also suggest two possible circuit-level implementations of clock stretching using clock gating and handshaking, respectively. Simulations on MCNC benchmark circuits with BPTM 70nm devices show that the proposed technique can achieve good yield by tolerating increased path delays (under variations and bridging defects of various sizes) with small overhead in performance and ~14% die-area compared to the conventional design. For performance analysis, we have implemented the proposed methodology in simple in-order pipeline in Simplescalar. Simulation results on SPEC2000 benchmarks show less that 2% of IPC (instructions-per-cycle) degradation.