Name
Abstract | ||
|---|---|---|
In this paper, a General Vector Machine (GVM) approximate calculation system based on FPGA is introduced. The author designed a parallel computing architecture of GVM on FPGA, discussed a matrix parallel computing method, developed approximation calculation methods of sigmoid function and softmax layer on FPGA. As an example, the paper implemented GVM on FPGA to identify MNIST data sets and tested the recognition rate with 14 different data accuracy of parameters, then gave some suggestions for data accuracy selection for parameters. Finally, the approximate calculation system is implemented and tested on XCKU3P, XC7Z020, XC7VX690 and XCUV190 FPGA chips. The results demonstrate that the computation speed is 112 times higher than CPU (Intel Core(TM) i9-7900X), and the performance can be equivalent to the GPU (NVIDIA GTX 1080 Ti). The approximate calculation system can effectively accelerate the calculation of GVM. It has the characteristics of good acceleration, and is suitable for embedded intelligent devices. |
Year | DOI | Venue |
|---|---|---|
2019 | 10.1016/j.mejo.2019.02.018 | Microelectronics Journal |
Keywords | Field | DocType |
GVM,Neural network accelerator,Parallel multiplication,Approximate calculation,FPGA | MNIST database,Softmax function,Matrix (mathematics),Support vector machine,Field-programmable gate array,Electronic engineering,Computational science,Acceleration,Engineering,Computation,Sigmoid function | Journal |
Volume | ISSN | Citations |
86 | 0026-2692 | 0 |
PageRank | References | Authors |
0.34 | 0 | 7 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| xuhui yang | 1 | 1 | 1.03 |
| Qingguo Zhou | 2 | 103 | 29.48 |
| Jinqiang Wang | 3 | 4 | 1.11 |
| Lihong Han | 4 | 0 | 1.01 |
| Fang Feng | 5 | 6 | 1.51 |
| Rui Zhou | 6 | 20 | 6.92 |
| Kuan-ching Li | 7 | 933 | 122.44 |