| dc.description.abstract | With the continuous and aggressive technology scaling, the design of memory systems becomes very challenging. The desire to have high-capacity, reliable, and energy efficient memory arrays is rising rapidly. However, from the technology side, the increasing leakage power and the restrictions resulting from the manufacturing limitations complicate the design of memory systems. In addition to this, with the new machine learning applications, which require tremendous amount of mathematical operations to be completed in a timely manner, the interest in neuromorphic systems has increased in recent years. Emerging Non- Volatile Memory (NVM) devices have been suggested to be incorporated in the design of memory arrays due to their small size and their ability to reduce leakage power since they can retain their data even in the absence of power supply.
Compared to other novel NVM devices, the Resistive Random Access Memory (RRAM) device has many advantages including its low-programming requirements, the large ratio between its high and low resistive states, and its compatibility with the Complementary Metal Oxide Semiconductor (CMOS) fabrication process. RRAM device suffers from other disadvantages including the instability in its switching dynamics and its sensitivity to process variations. Yet, one of the popular issues hindering the deployment of RRAM arrays in products are the RRAM reliability and radiation soft-errors. The RRAM reliability soft-errors result from the diffusion of oxygen vacations out of the conductive channels within the oxide material of the device. On the other hand, the radiation soft-errors are caused by the highly energetic cosmic rays incident on the junction of the MOS device used as a selector for the RRAM cell. Both of those soft-errors cause the unintentional change of the
resistive state of the RRAM device. While there is research work in literature to address some of the RRAM disadvantages such as the switching dynamic instability, there is no dedicated work discussing the impact of RRAM soft-errors on the various designs to which the RRAM device is integrated and how the soft-errors can be automatically detected and
fixed.
In this thesis, we bring the attention to the need of considering the RRAM soft-errors to avoid the degradation in design performance. In addition to this, using previously reported SPICE models, which were experimentally verified, and widely adapted system level simulators and test benches, various solutions are provided to automatically detect and
fix the degradation in design performance due to the RRAM soft-errors. The main focus in this work is to propose methodologies which solve or improve the robustness of memory systems to the RRAM soft-errors. These memories are expected to be incorporated in the current and futuristic platforms running the advanced machine learning applications. In
more details, the main contributions of this thesis can be summarized as:
- Provide in depth analysis of the impact of RRAM soft-errors on the performance of RRAM-based designs.
- Provide a new SRAM cell which uses the RRAM device to reduce the SRAM leakage power with minimal impact on its read and write operations. This new SRAM cell can be incorporated in the Graphical Processing Unit (GPU) design used currently
in the implementation of the machine learning platforms.
- Provide a circuit and system solutions to resolve the reliability and radiation soft-errors in the RRAM arrays. These solution can automatically detect and fix the soft-errors with minimum impact on the delay and energy consumption of the memory
array.
- A framework is developed to estimate the effect of RRAM soft-errors on the performance of RRAM-based neuromorphic systems. This actually provides, for the first time, a very generic methodology through which the device level RRAM soft-errors
are mapped to the overall performance of the neuromorphic systems. Our analysis show that the accuracy of the RRAM-based neuromorphic system can degrade by more than 48% due to RRAM soft-errors.
- Two algorithms are provided to automatically detect and restore the degradation in RRAM-based neuromorphic systems due to RRAM soft-errors. The system and circuit level techniques to implement these algorithms are also explained in this work.
In conclusion, this work offers initial steps for enabling the usage of RRAM devices in products by tackling one of its most known challenges: RRAM reliability and radiation soft-errors. Despite using experimentally verified SPICE models and widely popular system simulators and test benches, the provided solutions in this thesis need to be verified
in the future work through fabrication to study the impact of other RRAM technology shortcomings including: a) the instability in its switching dynamics due to the stochastic nature of oxygen vacancies movement, and b) its sensitivity to process variations. | en |
