Design of a Nonvolatile-Neural-Network-Accelerator-Embedded Edge-IoT Device and Its Hardware Emulation

Edge-IoT device with AI accelerator has advantages such as increased processing speed and more robust security, however, it has disadvantages such as limited resources and high power consumption during standby state caused by leakage current to retain internal data. To solve this issue, a nonvolatile-logic-based edge-IoT device which includes not only a nonvolatile AI accelerator but also a power management unit (PMU) and CPU is proposed and its functional behavior is emulated on a field-programmable gate array (FPGA). The use of the nonvolatile AI accelerator makes it possible to shorten the inference time and immediately fall into a power-off state without losing internal data, which results in great energy saving. The combination of the CPU with the PMU makes it possible to perform software-level management of the operation and power supply for the AI accelerator. As a typical design example, the proposed edge-IoT device which includes a binary convolutional neural network (BCNN) accelerator for digit recognition is designed and its function is emulated on a Digilent Basys3 FPGA board. The estimation result shows that the total energy for the periodical inference (1 image per 100μs) of the proposed edge-IoT device is reduced by 92% compared to that of a volatile edge-IoT device with a slight hardware overhead.