Side-channel attacks pose a serious security threat to embedded cryptographic devices. However, most of existing studies consider mono-channel attacks that make use of the leakage from a single channel. Despite recent studies confirming the effectiveness of non-profiled multi-channel attacks (MCAs) by combining the leakages from multiple channels, whether profiled multi-channel attack can be efficient remains unknown. In this work, we examine the performance of deep learning-based MCAs (DL-MCAs) while considering a wider range of targets including both unprotected and protected AES implementations. By employing the Gradient-weighted Class Activation Mapping (Grad-CAM) into MCAs, we contribute to the interpretablity of the performance of DL-MCAs. Meanwhile, we also propose a novel metric based on the Grad-CAM, by which the performance of DL-MCAs can be predicted during the training phase. Moreover, we demonstrate two special scenarios that DL-MCAs are very likely to outperform either mono-channel attack, and verify by attacks on simulated traces. Finally, the experimental results are turned out to be in accordance with the prediction using the proposed metric.