Non-volatile in-memory computing

The analysis of big-data at exa-scale (1018 bytes or flops) has called for an urgent need to re-examine the existing hardware platform that can support intensive data-oriented computing. A big-data-driven application requires huge bandwidth and yet able to ensure low-power density. For example, web-searching application involves crawling, comparing, ranking, and paging of billions of web-pages with extensive memory access. The existing memory technologies have critical challenges of scaling at nano-scale due to process variation, leakage current and I/O access limitations. Recently, the emerging non-volatile memory (NVM) technologies such as resistive-RAM (ReRAM), spin-transfer torque RAM (STT-RAM), domain-wall nanowire racetrack memory etc., have all shown significantly reduced standby power and increased integration density, not forgetting the close-to DRAM/SRAM access speed. Therefore, they are considered as promising candidates of universal memory for future big-data applications. The primary challenge to validate a hybrid design with both CMOS and nonvolatile devices is the lack of design platform that can validate the large-scale NVM circuit and system design accurately and efficiently. In addition, due to the use of non-electrical states of emerging NVM devices, new cells structures and their agreeing circuits for both read and write operations are needed to harness non-volatile memory with unique operations. For example, the transistor-free crossbar array that associates with NVM is different from conventional access transistor based memory structure. What is more, leveraging the NVM for computing, one also needs to examine the potential logic-in-memory computing architecture with significantly improved bandwidth and reduced power. In order to tackle above challenges ranging from device to system levels, this PhD thesis has explored the development of NVM design platform to support designs of non-volatile memories, readout and logic circuit designs, as well as the in-memory computing architecture. For the NVM design platform, the target is to perform accurate yet efficient circuit level simulation. The previous approaches either ignore dynamic effect without considering non-volatile states for dynamic behavior, or need equivalent circuits with high complexity to curve-fit non-linearity of those devices. We proposed a SPICE simulator named NVM-SPICE. This tool takes advantages of its new modified nodal analysis (MNA) framework, which can effectively support the non-electrical state variables of emerging non-volatile devices, such as ReRAM and spintronics devices. Due to the physics based modeling approach, NVM-SPICE is able to perform hybrid NVM/CMOS circuits efficiently and accurately. Compared to the equivalent circuit model based approach, the NVM-SPICE simulator exhibits more than 117x faster simulation speed for spintronics category devices and 40x faster speed for RRAM category devices. For NVM in-memory architecture, both memory elements and logic elements are implemented by emerging spintronics devices, which leads to a system purely composed of non-volatile devices. The detailed non-volatile memory and logic circuits are explored within the NVM-SPICE platform. In addition, logic is built inside the memory so that the I/O workload can be alleviated. Applications such as data retention, encryption, machine learning that play critical roles for big-data computing are explored within the non-volatile in-memory architecture. The evaluation results show that the purely non-volatile memory based platforms with in-memory architecture greatly contribute to power efficiency and throughput improvement for big-data oriented applications, and thus are potential candidates to be next generation information and communication technology.DOCTOR OF PHILOSOPHY (EEE