Boyang Chen, Mohd Tasleem Khan, George Goussetis, Mathini Sellathurai, Yuan Ding, João F. C. Mota, Jongeun Lee
COMET is a co-optimization framework for deploying Convolutional Neural Networks (CNNs) on resource-constrained FPGAs without any DSP blocks. It replaces conventional multiply-accumulate (MAC) units with a distributed arithmetic approach based on Offset-Binary Coding (OBC), combined with an im2col-based General Matrix Multiplication (GEMM) core.
The framework introduces two OBC schemes:
- Scheme A — OBC applied to inputs
- Scheme B — OBC applied to weights
This enables exploitation of bit-width asymmetry between inputs and weights for flexible hardware trade-offs.
- OBC-GEMM formulation of 2D CNN convolution layers using im2col, supporting both Scheme A and Scheme B
- Four LUT architectures — Parallel, Shared, Split, and Hybrid — with co-optimization analysis across resource usage, critical path delay, and power
- SA unit optimization unifying offset and bias terms into a single adder, reducing hardware overhead
- Quantization-aware training (QAT) evaluation on LeNet-5 (MNIST) and All-CNN-C (CIFAR-10) under fixed-point arithmetic
- FPGA evaluation on Xilinx ZCU106 and XCZU21DR, demonstrating competitive energy efficiency vs. state-of-the-art MAC-based accelerators
| Model | Platform | Freq | DSPs | Power | AEP |
|---|---|---|---|---|---|
| Modified LeNet-5 (HybridB) | ZCU106 | 95 MHz | 0 | 0.976 W | 0.70 ops/cycle/kENS |
| All-CNN-C (HybridB, w/ BRAM) | ZCU106 | 100 MHz | 0 | 1.236 W | 0.19 ops/cycle/kENS |
The Hybrid LUT architecture consistently achieves the best balance of LUT usage, flip-flop count, and power across all evaluated configurations.
@article{chen2025comet,
title={COMET: Co-optimization of a CNN model using efficient-hardware OBC techniques},
author={Chen, Boyang and Khan, Mohd Tasleem and Goussetis, George and Sellathurai, Mathini and Ding, Yuan and Mota, Jo{\~a}o FC and Lee, Jongeun},
journal={arXiv preprint arXiv:2510.03516},
year={2025}
}