Device Characteristic-Aware Quantization for eFlash-Based In-Memory Computing SoC

This study presents an eFlash-based in-memory computing (IMC) SoC leveraging device characteristic-aware quantization to enhance accuracy. A non-uniform quantization method, based on the non-linear characteristics of eFlash devices, is proposed to simplify the control of the number of electrons. Additionally, several channel-wise quantization schemes are introduced, considering the characteristics of the array and data distribution. These methods significantly improve image classification accuracy on the CIFAR-10 dataset using a CNN network, achieving enhancements of $\mathbf{2. 2 \%}, \mathbf{4. 2 \%}, \mathbf{5. 6 \%}$, and $\mathbf{1 4. 4} \%$ compared to baseline models.