STT-MRAM Co-design Deep Learning Model for IoT Applications

Machine/deep learning is one of key technologies to enhance the effectiveness of applications (e.g., vision and pattern recognition) for the Internet of Things (IoTs). Researchers strive to design an efficient structure using machine/deep learning in software-aided methods to accelerate computation. However, software-aided methods would be restricted in hardware resources such as capability of memory access. Memory, such as Static Random Access Memory (SRAM) and Spin-Transfer Torque Magnetic Random Access Memory (STT-MRAM), is critical component to effect the speed of data computation or access on applications. In this paper, we propose a hardware-aided model, STT-MRAM co-design deep learning model, to speed up such as vision recognition for the domain of IoTs, in which a STT-MRAM control circuit is designed to efficiently access data from memories and a software-based CPU is applied in a FPGA to process vision recognition. Notably, STT-MRAM is an emerging memory in lower power consumption, higher intelligence and non-volatile, which are critical for the design of IoT platform.