The proliferation of AI-enabled personal computers with heterogeneous processing units (CPU, GPU, NPU) introduces substantial complexity into resource scheduling due to dynamic neural network topologies that vary across inference phases and model architectures. This paper proposes a topology-aware adaptive scheduling algorithm that integrates real-time computational graph analysis into heterogeneous resource coordination mechanisms. The algorithm comprises three synergistic components: a lightweight runtime topology extraction module that captures evolving neural network structures, a predictive resource modeling system that forecasts device availability patterns, and an adaptive scheduling optimizer that jointly considers topology dependencies, device heterogeneity, and temporal resource fluctuations. Experimental evaluation across six representative neural architectures (ResNet-50, MobileNetV3, YOLOv8, BERT-Base, Vision Transformer, EfficientNet-B4) and three heterogeneous AI-PC platforms demonstrates substantial performance improvements: 13.5% latency reduction, 15.6% throughput increase, 30.1% NPU utilization gain, and 16.1% energy efficiency improvement compared to state-of-the-art baselines. The algorithm exhibits robust performance under dynamic workload conditions and diverse topological structures, maintaining functional operation even under resource failures and measurement uncertainties.
Shao et al. (Mon,) studied this question.