一种面向神经网络在FPGA上高效部署的混合精度自动化优化方法
By constructing a hardware resource consumption prediction model and using adaptive penalized augmented Lagrange optimization, the problems of low efficiency and poor reliability in neural network deployment on FPGA are solved, and efficient mixed-precision configuration and deployment under hardware constraints are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF SCI & TECH
- Filing Date
- 2026-05-26
- Publication Date
- 2026-07-17
AI Technical Summary
Existing methods for deploying neural networks on FPGAs suffer from problems such as low optimization efficiency, lack of hardware feature awareness, and reliance on labeled data, making it difficult to achieve the optimal balance between accuracy and hardware overhead.
A hardware resource consumption prediction model is constructed, a graph-structured perceptual network is used for accurate configuration prediction, and an adaptive penalized augmented Lagrange method is used for constraint optimization. Combined with hierarchical neighborhood search, adapted hardware code is generated to meet FPGA resource and timing constraints.
It enables efficient deployment of neural networks on FPGAs, significantly improving inference energy efficiency and deployment efficiency, reducing dependence on labeled data, and expanding applicable scenarios.
Smart Images
Figure CN122261673B_ABST