The 8-bit floating-point (FP8) format is promising for FPGA-based DNN acceleration because they offer a wider dynamic range than integer formats with similar storage cost. This project explores Log-Mul, a high-accuracy approximate multiplier for FP8 computation on FPGAs. Log-Mul uses a hardware-friendly logarithmic approximation to simplify multiplication while supporting multiple FP formats. We are currently developing a FPGA-based architecture for Log-Mul and integrating it into a diffusion model accelerator as a case study. Preliminary evaluation on AMD UltraScale+ FPGA focuses on numerical error, model accuracy, resource utilization, and critical-path delay. This work aims to demonstrate approximate low-bitwidth floating-point multiplication as a practical design option for efficient FPGA-based DNN acceleration.
Chen, R, Wang, X, Stroobandt, D & da Silva, B 2026, High-Accuracy FPGA-Based FP8 Approximate Multiplier. in 2026 36th International Conference on Field-Programmable Logic and Applications (FPL). pp. 374-375. https://doi.org/10.23919/FPLCPS00026.2026.00072
Chen, R., Wang, X., Stroobandt, D., & da Silva, B. (2026). High-Accuracy FPGA-Based FP8 Approximate Multiplier. In 2026 36th International Conference on Field-Programmable Logic and Applications (FPL) (pp. 374-375) https://doi.org/10.23919/FPLCPS00026.2026.00072
@inproceedings{f6f958638feb4ec98f2f24533904077c,
title = "High-Accuracy FPGA-Based FP8 Approximate Multiplier",
abstract = "The 8-bit floating-point (FP8) format is promising for FPGA-based DNN acceleration because they offer a wider dynamic range than integer formats with similar storage cost. This project explores Log-Mul, a high-accuracy approximate multiplier for FP8 computation on FPGAs. Log-Mul uses a hardware-friendly logarithmic approximation to simplify multiplication while supporting multiple FP formats. We are currently developing a FPGA-based architecture for Log-Mul and integrating it into a diffusion model accelerator as a case study. Preliminary evaluation on AMD UltraScale+ FPGA focuses on numerical error, model accuracy, resource utilization, and critical-path delay. This work aims to demonstrate approximate low-bitwidth floating-point multiplication as a practical design option for efficient FPGA-based DNN acceleration. ",
author = "Ruiqi Chen and Xiaoke Wang and Dirk Stroobandt and \{da Silva\}, Bruno",
year = "2026",
month = oct,
day = "27",
doi = "10.23919/FPLCPS00026.2026.00072",
language = "English",
pages = "374--375",
booktitle = "2026 36th International Conference on Field-Programmable Logic and Applications (FPL)",
}