ParallelFusion: Towards Maximum Utilization of Mobile GPU for DNN Inference

Mobile GPUs are extremely under-utilized for DNN computations across different mobile deep learning frameworks and multiple DNNs with various complexities. We explore the feasibility of batching and it improves the throughput by up to 35%. However, real-time applications in mobile have a limited amount of requests to get a benefit from batching. To tackle the challenge, we present ParallelFusion technique that enables concurrent execution of heterogeneous operators to further utilize the mobile GPU. We implemented ParallelFusion over the MNN framework and evaluated on 6 state-of-the-art DNNs. Our evaluation shows that Parallel Fusion achieves up to 195% to 218% throughput with fused execution of 2 and 3 operators compared to single DNN inference.