Applications of Emerging Memory in Modern Computer Systems: Storage and Acceleration

In recent year, heterogeneous architecture emerges as a promising technology to conquer the constraints in homogeneous multi-core architecture, such as supply voltage scaling, off-chip communication bandwidth, and application parallelism. Various forms of accelerators, e.g., GPU and ASIC, have been extensively studied for their tradeoffs between computation efficiency and adaptivity. But with the increasing demand of the capacity and the technology scaling, accelerators also face limitations on cost-efficiency due to the use of traditional memory technologies and architecture design. Emerging memory has become a promising memory technology to inspire some new designs by replacing traditional memory technologies in modern computer system. In this dissertation, I will first summarize my research on the application of Spin-transfer torque random access memory (STT-RAM) in GPU memory hierarchy, which offers simple cell structure and non-volatility to enable much smaller cell area than SRAM and almost zero standby power. Then I will introduce my research about memristor implementation as the computation component in the neuromorphic computing accelerator, which has the similarity between the programmable resistance state of memristors and the variable synaptic strengths of biological synapses to simplify the realization of neural network model. At last, a dedicated interconnection network design for multicore neuromorphic computing system will be presented to reduce the prominent average latency and power consumption brought by NoC in a large size neuromorphic computing system