Low-Power Approximate Multipliers Based on Novel 4-2 and 5-2 Compressors for CNN Inference

초록

Recent approximate computing and multiplier studies exploit application/data distributions to reduce arithmetichardware cost. This paper proposes new 4-2 and 5-2 approximate compressors and C-FULL 8×8 Daddamultipliers. In MNIST CNN inference, the proposed hybrid multiplier maintains 97.3% accuracy, comparable to theexact multiplier’s 97.2%, while reducing PPA by 76% and normalized FF/LUT by 6.8%/28.6%. Further evaluation onCIFAR-10 shows that 4-2 error correction mitigates the accuracy degradation of the proposed hybrid multipliers.

키워드