Skip to Main content Skip to Navigation

Développement d'architectures HW/SW tolérantes aux fautes et auto-calibrantes pour les technologies Intégrées 3D

Abstract : 3D technology promises energy-efficient heterogeneous integrated systems, which may open the way to thousands cores chips. Silicon dies containing processing elements are stacked and connected by vertical wires called Through-Silicon-Vias. In 3D chips, interconnecting an increasing number of processing elements requires a scalable high-performance interconnect solution: the 3D Network-on-Chip. Despite the advantages of 3D integration, testing, reliability and yield remain the major challenges for 3D NoC-based systems. In this thesis, the TSV interconnect test issue is addressed by an off-line Interconnect Built-In Self-Test (IBIST) strategy that detects both structural (i.e. opens, shorts) and parametric faults (i.e. delays and delay due to crosstalk). The IBIST circuitry implements a novel algorithm based on the aggressor-victim scenario and alleviates limitations of existing strategies. The proposed Kth-aggressor fault (KAF) model assumes that the aggressors of a victim TSV are neighboring wires within a distance given by the aggressor order K. Using this model, TSV interconnect tests of inter-die 3D NoC links may be performed for different aggressor order, reducing test times and circuitry complexity. In 3D NoCs, TSV permanent and transient faults can be mitigated at different abstraction levels. In this thesis, several error resilience schemes are proposed at data link and network levels. For transient faults, 3D NoC links can be protected using error correction codes (ECC) and retransmission schemes using error detection (Automatic Retransmission Query) and correction codes (i.e. Hybrid error correction and retransmission).For transients along a source-destination path, ECC codes can be implemented at network level (i.e. Network-level Forward Error Correction). Data link solutions also include TSV repair schemes for faults due to fabrication processes (i.e. TSV-Spare-and-Replace and Configurable Serial Links) and aging (i.e. Interconnect Built-In Self-Repair and Adaptive Serialization) defects. At network-level, the faulty inter-die links of 3D mesh NoCs are repaired by implementing a TSV fault-tolerant routing algorithm. Although single-level solutions can achieve the desired yield / reliability targets, error mitigation can be realized by a combination of approaches at several abstraction levels. To this end, multi-level error resilience strategies have been proposed. Experimental results show that there are cases where this multi-layer strategy pays-off both in terms of cost and performance. Unfortunately, one-fits-all solution does not exist, as each strategy has its advantages and limitations. For system designers, it is very difficult to assess early in the design stages the costs and the impact on performance of error resilience. Therefore, an error resilience exploration (ERX) methodology is proposed for 3D NoCs.
Document type :
Complete list of metadatas
Contributor : Abes Star :  Contact
Submitted on : Wednesday, January 22, 2014 - 3:37:20 PM
Last modification on : Friday, December 11, 2020 - 8:28:05 AM
Long-term archiving on: : Wednesday, April 23, 2014 - 6:45:41 AM


Version validated by the jury (STAR)


  • HAL Id : tel-00838677, version 2




Vladimir Pasca. Développement d'architectures HW/SW tolérantes aux fautes et auto-calibrantes pour les technologies Intégrées 3D. Autre. Université de Grenoble, 2013. Français. ⟨NNT : 2013GRENT001⟩. ⟨tel-00838677v2⟩



Record views


Files downloads