Similar Items: DPU or GPU for Accelerating Neural Networks Inference -- Why not both? Split CNN Inference