In this paper, an architecture is proposed for energy-efficient DNN training. It leverages triple sparsity in training which eliminates more unnecessary computation compared with prior works. Meanwhile, a 2-level grained mask-matching scheme is introduced for efficient sparsity detection. Coarse-grained mask match units are reused among PEs to save power and area while fine-grained mask match units are inside each PE to maintain throughput. As a result, our architecture achieves 42.1 TOPS and 174.0 TOPS/W in terms of throughput and energy efficiency, respectively. The energy efficiency of our design is 2.12× higher than the state-of-the-art training processor.