Data and Fault Aware Routing Algorithm for NoC Based Approximate Computing

Due to transistor shrinking and core number increasing in System-on-Chip (SoC), fault tolerance has become a critical concern. Given the amount of data communications on such architectures, Network-on-Chips (NoCs) lead a crucial role in terms of performance. Even if fault correction approaches have been developed, they cannot efficiently address several permanent faults on NoC, due to their high hardware costs and correction limitations. In parallel, Approximate Computing domain considers applications that can tolerate errors, hence allowing fault mitigation instead of correction. This latter brings the opportunity of low implementation cost techniques to improve the reliability of SoC. In this work, we propose a routing technique which selects a path between cores according to data type and permanent fault positions. Error tolerant data are able to cross faulty paths by using a bit-shuffling error mitigation technique. Critical data circumvent faulty paths or are duplicated and shuffled in case there is no other correct path available. Results show that our routing technique allows to maintain all the communication paths within the NoC for a large amount of permanent errors. To further evaluate the behavior of the proposed technique, we performed a comprehensive analysis of the technique on the packet latency and saturation injection rate with respect to the number of faults and traffic type.