Fault Tolerance of Stochastic Decoders for Error Correcting Codes

Abstract

Low-density Parity-check (LDPC) codes are very powerful linear error-correcting codes, first introduced by Gallager in 1963. They are now used in many communication standards due to their ability to achieve near Shannon-capacity performance. Stochastic decoding is a hardware-efficient method of iterative decoding of LDPC codes. In this work, we investigate the capability of stochastic decoding to tolerate circuit soft errors while maintaining good bit error rate performance and low error floor. Soft errors can be intended faults as a result of either supply voltage scaling to reduce power consumption or overclocking the system to achieve a higher throughput. They can also be unintended faults as a result of temperature or process variations.

We develop two models to emulate these circuit errors at the system level. We apply our models to two standardized LDPC codes (10GBASE-T and WiMAX). Simulation results show that stochastic decoding is very tolerant to faults and errors, where it can tolerate a probability of setup time violation of 0.1 in the wires of the decoder. Hence, stochastic decoding can be very useful in systems with very low power or high performance requirements where we can push the limits of power or speed by lowering the supply voltage or highly overclocking the system while maintaining good performance. In addition, a chip has been designed and sent to fabrication to do post-silicon validation and verify our models.