SparOA: Sparse and Operator-aware Hybrid Scheduling for Edge DNN Inference

TL;DR

SparOA tackles real-time edge DNN inference by modeling per-operator sparsity $\rho$ and computational intensity $I$ as orthogonal axes for CPU-GPU scheduling. It combines a Transformer-LSTM threshold predictor with an SAC-based operator scheduler to adaptively allocate work between CPU and GPU, while an asynchronous hybrid engine with dynamic batching hides data-transfer latency. The approach yields substantial latency improvements (average $1.22\times$ to $1.31\times$ over SOTA and up to $50.7\times$ over CPU-Only) and reduced energy-per-inference ($7\%$–$16\%$ lower than the SOTA co-execution baseline). These results on NVIDIA Jetson platforms demonstrate SparOA's practical viability for diverse edge models including CNNs and transformers, and highlight the benefits of joint sparsity/intensity analysis with RL-driven scheduling.

Abstract

The resource demands of deep neural network (DNN) models introduce significant performance challenges, especially when deployed on resource-constrained edge devices. Existing solutions like model compression often sacrifice accuracy, while specialized hardware remains costly and inflexible. Hybrid inference methods, however, typically overlook how operator characteristics impact performance. In this work, we present SparOA, a CPU-GPU hybrid inference framework, which leverages both sparsity and computational intensity to optimize operator scheduling. SparOA embraces aforementioned challenges through three key components: (1) a threshold predictor that accurately determines optimal sparsity and computational intensity thresholds; (2) a reinforcement learning-based scheduler that dynamically optimizes resource allocation based on real-time hardware states; and (3) a hybrid inference engine that enhances efficiency through asynchronous execution and batch size optimization.Extensive results show that SparOA achieves an average speedup of 1.22-1.31x compared to all baselines, and outperforms the CPU-Only by up to 50.7x. Also, SparOA achieves optimal energy-per-inference, consuming 7\%-16\% less energy than the SOTA co-execution baseline.

Figures (12)

• Figure 1: Overview of SparOA.
• Figure 2: The distribution of sparsity and computational intensity for each operator in MobileNetV3-small, measured on NVIDIA Jetson AGX Orin. We set the batch size to 1.
• Figure 3: The architecture of threshold predictor.
• Figure 4: Example showing how SparOA calculates different operators.
• Figure 5: Inference latency comparison of SparOA against the baselines, depending on target device.