2019 2nd International Conference on Advanced Materials, Intelligent Manufacturing and Automation | |
An FPGA-Based CNN Efficient Storage Processor | |
Zhao, Tong^1 ; Qiao, Lufeng^1 ; Chen, Qinghua^1 | |
Institute of Communication Engineering, Army Engineering University of PLA, Nanjing Jiangsu | |
210007, China^1 | |
关键词: Bandwidth requirement; Computational performance; Winograd; | |
Others : https://iopscience.iop.org/article/10.1088/1757-899X/569/3/032013/pdf DOI : 10.1088/1757-899X/569/3/032013 |
|
来源: IOP | |
【 摘 要 】
At present, deep learning algorithms such as neural networks are widely used in various aspects of artificial intelligence. The computational performance of the CPU is low and the power consumption of the GPU is large. Therefore, this paper studies the implementation of CNN on the FPGA and proposes an effective storage management scheme, which greatly reduces the bandwidth requirement of the operation. The Winograd algorithm is used in the operation to reduce the computational complexity of the convolution, so that the performance of the FPGA is optimized. This design implements Alexnet on the Virtex7 xc7vx690t with 1.32 TFlop/s performance and an average Energy efficiency of 43.5 GOP/s/W.
【 预 览 】
Files | Size | Format | View |
---|---|---|---|
An FPGA-Based CNN Efficient Storage Processor | 817KB | download |