3D pose estimation is a core step for human-computer interaction and human action recognition. However, time-sensitive applications like virtual reality also need this task to achieve real-time speed. This paper proposes a multitask and multi-level neural network architecture with a highspeed friendly 3D human pose representation. Based on this, we build a real-time multi-person 3D pose estimation system with a single RGB image as input. The network estimates 3D poses from the input image directly by the multi-task design and keeps both accuracy and speed by the multi-level detection design. By evaluation, we show our system achieves the 21 fps on RTX 2080 with only 33 mm accuracy lose compared with related works. We also provide network visualization to prove our network work as we design. This work shows the possibility for a single RGB image based 3D pose estimation system to achieve real-time speed, which is a basement for building a low-cost 3D motion capture system.