Efficient FPGA Sigmoid Implementation via Mixed-Radix CORDIC
This study introduces an FPGA implementation of the sigmoid activation function that is efficient in terms of hardware, utilizing a mixed-radix CORDIC architecture. By exploiting the connection between sigmoid and hyperbolic tangent functions, the input range is normalized to 1, facilitating tanh calculations within a limited range of 0.5, which enhances convergence. A refined mixed-radix hyperbolic rotation CORDIC provides high precision while maintaining low hardware demands, making it suitable for deployment in resource-limited edge devices such as FPGAs for neural networks.
Key facts
- Efficient hardware implementation of nonlinear activation functions is crucial for deploying ANNs on FPGAs.
- Sigmoid is widely used for probabilistic output, binary classification, and gating in RNNs.
- The proposed approach uses a mixed-radix CORDIC-based architecture.
- Input range is normalized to 1, enabling tanh computation within a reduced range of 0.5.
- A modified mixed-radix hyperbolic rotation CORDIC achieves high accuracy with minimal hardware overhead.
- The implementation targets resource-constrained and edge devices such as FPGAs.
- The paper is available on arXiv with ID 2604.23547.
- The method leverages the mathematical relationship between sigmoid and hyperbolic tangent functions.
Entities
Institutions
- arXiv