Improving Phase Change Memory (PCM) and Spin-Torque-Transfer Magnetic-RAM (STT-MRAM) as Next-Generation Memories: A Circuit Perspective

In the memory hierarchy of computer systems, the traditional semiconductor memories Static RAM (SRAM) and Dynamic RAM (DRAM) have already served for several decades as cache and main memory. With technology scaling, they face increasingly intractable challenges like power, density, reliability and scalability. As a result, they become less appealing in the multi/many-core era with ever increasing size and memory-intensity of working sets.\ud \ud Recently, there is an increasing interest in using emerging non-volatile memory technologies in replacement of SRAM and DRAM, due to their advantages like non-volatility, high device density, near-zero cell leakage and resilience to soft errors. Among several new memory technologies, Phase Change Memory (PCM) and Spin-Torque-Transfer Magnetic-RAM (STT-MRAM) are most promising candidates in building main memory and cache, respectively. However, both of them possess unique limitations that preventing them from being effectively adopted.\ud \ud In this dissertation, I present my circuit design work on tackling the limitations of PCM and STT-MRAM. At bit level, both PCM and STT-MRAM suffer from excessive write energy, and PCM has very limited write endurance. For PCM, I implement Differential Write to remove large number of unnecessary bit-writes that do not alter the stored data. It is then extended to STT-MRAM as Early Write Termination, with specific optimizations to eliminate the overhead of pre-write read. At array level, PCM enjoys high density but could not provide competitive throughput due to its long write latency and limited number of read/write circuits. I propose a Pseudo-Multi-Port Bank design to exploit intra-bank parallelism by recycling and reusing shared peripheral circuits between accesses in a time-multiplexed manner. On the other hand, although STT-MRAM features satisfactory throughput, its conventional array architecture is constrained on density and scalability by the pitch of the per-column bitline pair. I propose a Common-Source-Line Array architecture which uses a shared source-line along the row, essentially leaving only one bitline per column.\ud \ud For these techniques, I provide circuit level analyses as well as architecture/system level and/or process/device level discussions. In addition, relevant background and work are thoroughly surveyed and potential future research topics are discussed, offering insights and prospects of these next-generation memories