The widespread adoption of Deep Neural Networks (DNNs) across diverse applications has intensified the demand for efficient execution strategies on GPU-CPU heterogeneous devices. While existing device-oriented DNN partitioning methods optimize either energy consumption or execution latency in isolation, they often lack the capability to dynamically adapt partitioning strategies under real-world constraints. To address these limitations, we first formulate the DNN partitioning problem as an optimal execution path search on a directed acyclic graph (DAG), then propose an imitation learning (IL)-based DNN partitioning framework that dynamically allocates DNN layers across GPU-CPU processors to minimize energy consumption while satisfying time constraints. Specifically, we first design aProfiler, which analyzes the layer-wise characteristics such as the execution time and energy consumption of layers in DNNs executed on the GPU-CPU heterogeneous device. Then, we present anOracle Generatorthat generates high-quality partitioning solutions based on the Dijkstra algorithm to form a training datasetOracle.Oraclefrom the proposed generator is input to a proposed Trainer, which iteratively learns from theOracledata to obtain a regression model namedPredictor.Predictorcan predict the optimal partitioning strategy for DNNs in real time. Finally, we design aData Aggregator, which enhancesOracledata through continuous runtime feedback, thereby improving the prediction accuracy of the regression model. We evaluate the proposed IL-based DNN partitioning method on 3 NVIDIA Jetson platforms and 1 Huawei NPU platform. Experimental results demonstrate that compared to 3 benchmarking methods, our method reduces energy consumption by up to 73.15%.

This letter focuses on integrated CPU-GPU edge systems with renewable energy sources and studies the resource management problem to minimize the energy consumption of real-time tasks while ensuring temperature and reliability constraints. We propose an imitation learning (IL)-based resource management scheme, ILRM, implemented in two phases: 1) offline Oracle generation and 2) online IL. In the offline phase, we design a fast-converging heuristic to generate near-optimal solutions (i.e., Oracles) for training an online prediction model. In the online phase, we realize IL using the trained model that predicts the resource configuration policies for the incoming task sets to be scheduled. A data aggregation method is also developed to enhance the robustness of the prediction model. We validate ILRM through extensive experiments on both simulated and real integrated CPU-GPU edge platforms.