OCEAN: An on-chip incremental-learning enhanced processor with gated recurrent neural network accelerators

Abstract

A deep learning processor with 8 gated recurrent neural network (RNN) accelerators is proposed in this paper. It features on-chip incremental learning by numerical and local gradient computation enhancement. Extra precision of training is obtained without extending the bit-width. Tri-mode weight access (DMA/FIFO/RAM) improves the throughput during incremental learning. The number multipliers and activation function engines are reduced by hardware sharing. The processor is fabricated in 65nm CMOS, consumes 155mW at 400MHz /1.2V or 6.6mW at 20MHz /0.8V. It achieves a peak throughput of 54.2Kfps and energy efficiency of 0.24μJ/classification on MNIST. Demonstrations are accomplished on sequential AI tasks such as text-based motion detection and real-time wake-up word speech recognition.