Predictive Modeling of FPGA Resource and Power Consumption for Configurable CNN Operators

Philippe Magalhães, Virginie Fresse, Benoît Suffran, Olivier AlataICECS 2025 · DOI

Deploying Convolutional Neural Networks (CNNs) on FPGAs presents a significant design challenge: balancing high-performance parallelism with finite hardware resources. Traditionally, designers rely on a "trial and error" approach, where configurations are synthesized through tools like Xilinx Vivado to determine if they fit within the target device's logic, memory, and power constraints. This process can take hours, severely bottlenecking Design Space Exploration (DSE). This paper proposes a solution by developing a library of parameterizable VHDL components and a mathematical framework to predict hardware footprints before synthesis.

Methodology

The researchers developed a two-phase framework to bypass the synthesis bottleneck. First, they created a custom library of Intellectual Property (IP) blocks for core CNN operations, including convolution, activation (ReLU/ReLU6), and pooling. These blocks were designed with varying optimization targets, such as low logic usage or high throughput via Digital Signal Processors (DSPs).

Second, they established a modeling pipeline to characterize these IPs. By synthesizing hundreds of variations—varying data and coefficient bit-widths from 3 to 16 bits—they generated a robust dataset of hardware reports. Using Pearson’s correlation, they identified key parameters influencing resource consumption, followed by the application of polynomial and piecewise regression to derive predictive equations.

The proposed methodology for generating predictive mathematical models from synthesis data.

Modeling Architectural Shifts

A critical insight from the study is that simple linear models are often insufficient for hardware prediction. As IP configurations change, internal architectural shifts—such as transitioning from serial to parallel processing—cause abrupt "jumps" in resource usage.

To address these discontinuities, the authors utilized piecewise regression. This allows the models to capture the specific thresholds where resource consumption transitions, ensuring high accuracy even when the underlying hardware implementation changes significantly based on input parameters.

Resource usage (LLUTs) as a function of data and coefficient bits, demonstrating the need for piecewise modeling to capture architectural discontinuities.

Key Results and Validation

The predictive models demonstrated remarkable precision when tested on a Xilinx Zynq UltraScale+ ZCU104 platform:

  • High Accuracy: The models achieved a Coefficient of Determination ($R^2$) greater than 0.94 for Logic Look-Up Table (LLUT) predictions, with many reaching 0.99.
  • Low Error Rates: The Mean Absolute Percentage Error (MAPE) remained generally below 8%.
  • System-Level Performance: In a full 8-bit CNN architecture validation, the error for primary resources (LLUT, FF, DSP) was under 1.3%. Even complex metrics like Carry Chains (CChains) were predicted with a manageable 9.5% error.

Conclusion

By providing a way to estimate resource and power consumption in seconds rather than hours, this framework enables rapid Design Space Exploration. While currently optimized for the Xilinx UltraScale+ family, the methodology is tool-agnostic and can be adapted to other FPGA vendors. Future work aims to incorporate operating frequency and timing constraints into the models to further automate the "right-sizing" of CNNs for specific hardware targets.

Summary updated

BibTeX
@inproceedings{magalhaes2025predictive,
  author = {Philippe Magalhães and Virginie Fresse and Benoît Suffran and Olivier Alata},
  title = {{Predictive Modeling of FPGA Resource and Power Consumption for Configurable CNN Operators}},
  booktitle = {ICECS},
  year = {2025},
  doi = {10.1109/icecs66544.2025.11270578},
}
Figure