This paper presents OCEAN: an artificial neural network processor designed for accelerating gated-recurrent-unit (GRU) inference and on-chip incremental learning for sequential modeling. Implemented in 65-nm CMOS with silicon area of 2.9 × 3.5 mm 2 , the OCEAN processor features a 32-bit reduced instruction set computing core, 64-KB on-chip SRAM, and eight 16-bit four-cell GRU accelerators for inference and gradient computation. Each GRU accelerator is optimized and enhanced for efficient gradient computation. The processor is measured to consume 155 mW at the peak clock rate of 400 MHz and the supply of 1.2 V or 6.6 mW at 20 MHz/0.8 V. Both inference and on-chip incremental learning are accomplished on well-known AI tasks such as handwritten digit recognition, semantic natural language processing, and biomedical waveform-based seizure detection.
1 May 2019 in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
The paper presents iFPNA, instruction-and-fabric programmable neuron array: a deep learning processor using a neural network specific instruction set architecture and reconfigurable fabric. The design is motivated by the trade-off between the efficiency and flexibility of deep learning processor designs. The proposed architecture contains a controller for programming, a global feature buffer for data arrangement and 16 reconfigurable neuron slices for computing. The controller uses dedicated instructions to control the feature buffer and neuron slices. The neuron slices support multiplication-andaccumulation, non-linear activation, element-wise operation and pooling of different bit-width and kernel size. Different deep neural networks are mapped to iFPNA, such as conventional CNNs, depthwise/pointwise CNNs, LSTMs and GRUs. The paper also discusses various data flow mapping schemes. An iFPNA prototype is designed and fabricated on 28nm HPC CMOS technology. Measurement results show that the iFPNA achieves a peak energy efficiency of 1.72 TOPS/W running at 30 MHz clock rate with 0.63 V voltage supply. The measured latency on AlexNet is 57.2 ms and on LSTM-512 is 40 ms at 125 MHz clock rate.