Domain Specific Acceleration

In response to the increasing demand for deep learning computing power, the laboratory began to accelerate its research on deep neural networks in 2013. In 2016, we took the lead in proposing structured sparse pruning and dynamic low-bit quantization technology. With almost no loss of accuracy, the model calculation amount and bandwidth requirements are compressed by 10-20 times. The FPGA-specific accelerators designed are compared with CPUs and GPUs. Achieve 40 times and 16 times energy efficiency improvement. In 2017, it further proposed the "pruning-quantization-customization" neural network software and hardware co-design paradigm widely used in the industry, proposed coarse-grained instruction set architecture and layer fusion compilation technology, and designed FPGA-oriented deep learning processors and end-to-end deployment tools The chain reduces the deployment cost of any model to the order of hundreds of seconds, breaking through the problem of long development cycle of intelligent algorithm FPGA dedicated hardware. The research results were transformed into shares of Beijing Shenjian Technology Co., Ltd., and at the end of 2018, it was acquired by Xilinx, a leading company in reconfigurable computing, for about US$300 million, which promoted the industrialization of FPGA deep learning accelerators on a global scale.