In recent years, compressed sensing techniques have been applied to the reconstruction of parallel magnetic resonance (MR) images. Particularly for 3D MR signal, it is crucial to acquire fewer samples to reduce the distortions caused by long-time acquisitions (e.g., motion, organ dynamics). Motivated by the recent success of ConvNet in 2D image reconstruction, we propose to extend the approach to 3D volume reconstruction and parallel MR imaging. The structure of the proposed network follows FBPConvNet with additional coil compression by SSoS and wavelet transform. A parallelism using two GPUs is also applied to overcome the memory shortage. The proposed method is able to reconstruct a (320 x 320 x 256 x 8) volume in less than 10s with 2 GPUs, while the iterative algorithm l(1)-ESPIRiT takes over 5 min in CPU.