Accurate transient simulation of power semiconductors is essential for the design and validation of power electronic systems. However, a long-standing trade-off exists between model accuracy and computational efficiency, especially in real-time simulations. This paper proposes a super-sampling method that actively optimizes simulation output waveform accuracy without increasing model complexity. A hybrid strategy is adopted, where a five-layer Convolutional Neural Network (CNN) predicts high-resolution waveform values at sampling points, and the remaining points are recovered through linear interpolation. This allows the numerical model to be solved using larger simulation steps with lower computational burden, while the waveform accuracy is subsequently improved at a minimal cost. The proposed method is evaluated on an IGBT model. Experimental results show that the enhanced waveforms achieve an average relative root mean square error (RRMSE) below 1%, significantly outperforming direct interpolation methods. To further reduce errors in regions with rapid changes, a refined Upsample-CNN is introduced and validated. The successful application of the proposed method breaks through the traditional trade-off, offering a new perspective and promoting power electronics simulation techniques.