Improving Efficiency in Neural Network Accelerator Using Operands Hamming Distance optimization
This work addresses energy efficiency for on-device AI inference, offering a domain-specific improvement that is incremental in nature.
The paper tackled the problem of high data-path energy consumption in neural network accelerators by optimizing the Hamming distance of input operands, achieving an average 2.85x and up to 8.51x reduction in data-path energy for MobileNetV2.
Neural network accelerator is a key enabler for the on-device AI inference, for which energy efficiency is an important metric. The data-path energy, including the computation energy and the data movement energy among the arithmetic units, claims a significant part of the total accelerator energy. By revisiting the basic physics of the arithmetic logic circuits, we show that the data-path energy is highly correlated with the bit flips when streaming the input operands into the arithmetic units, defined as the hamming distance of the input operand matrices. Based on the insight, we propose a post-training optimization algorithm and a hamming-distance-aware training algorithm to co-design and co-optimize the accelerator and the network synergistically. The experimental results based on post-layout simulation with MobileNetV2 demonstrate on average 2.85X data-path energy reduction and up to 8.51X data-path energy reduction for certain layers.