Reliability and Robustness Enhancement of RRAM-based Neuromorphic Computing

• Applicants: Privatdozent Dr.-Ing. Bing Li; Professor Dr.-Ing. Ulf Schlichtmann; Professorin Dr.-Ing. Li Zhang
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; Electronic Semiconductors, Components and Circuits, Integrated Systems, Sensor Technology, Theoretical Electrical Engineering
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 457473137

Deep neural networks (DNNs) have successfully been applied to solve complex problems in various fields such as speech and image processing. In a DNN, a huge number of multiply-accumulate (MAC) operations are required. Implementing the MAC operations by Ohm’s law and Kirchhoff’s law in the analog domain, RRAM-based crossbars have exhibited a great potential in performance and energy efficiency for accelerating such computing operations. However, due to this analog nature, RRAM-based crossbars are affected severely by process variations, noise, non-ideal device and programming characteristics etc. Consequently, the computing accuracy of such crossbar accelerators degrades significantly if these effects are not addressed. In view of their great potential, RRAM-based crossbars have been investigated intensively in the research community in recent years. However, their robustness and reliability issues have not been addressed systematically. Therefore, in this project we propose to explore a systematic solution to deal with these issues using techniques ranging from complementary hardware redundancy to DNN architectural modification. Specifically, we will first model the hardware uncertainties of RRAM cells and crossbars statistically and incorporate them into a statistical training and inference engine. With this engine, we will be able to introduce and evaluate techniques such as error correction and filter expansion applied onto given DNN architectures to improve their robustness. Reliability issues such as aging and thermal effects will be countered proactively by weight redistribution and balancing. Furthermore, complementary hardware redundancy will be explored to realize a codesign scheme to strengthen the RRAM-based computing system. The resulting system will be verified by a simulation platform, which also provides corner-based low-level sign-off simulation to certify the complete system. Since this framework addresses general robustness and reliability issues in analog-based neuromorphic computing, it can also be adapted for other neuromorphic computing systems based on e.g. SRAM, Flash, or spintronic devices, and therefore benefit the whole landscape of deep learning and artificial intelligence.

• DFG Programme: Research Grants