A Fast and Low Power Hardware Accelerator for ANN Working at Near Threshold Voltage

摘要

Artificial neural network (ANN) are widely applied in machine learning and artificial intelligence. But ANN usually requires large data throughputs and induces high power consumptions. This paper proposes an accelerator design guideline for ANN with full consideration of the hardware scale, performance and power consumption. We apply multiple clocks in our design to get a high data throughput and introduce the near threshold voltage (NTV) to get a lower power consumption. We further optimize the multiplication operation in critical path obtaining a higher performance.