Budget-Constrained Multi-Stage Search for Deep Learning Deployment Configurations on Ascend 310B Edge NPU

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Deploying deep learning models on edge neural processing units (NPUs) requires selecting multiple compilation and runtime parameters. Exhaustive evaluation is costly, so high-quality configurations must be found within a limited hardware evaluation budget. We propose a budget-constrained, multi-stage search method for the Ascend 310B platform and evaluate it on an Ascend 310B1 device. An L9 orthogonal design first provides low-cost, structured measurements of the main effects of precision mode, batch size, operator implementation mode, and compilation and execution strategy. CRITIC assigns fixed objective weights to the measured performance metrics. These main effects then prioritize untested configurations for device-side validation, and TOPSIS ranks the feasible configurations. We measured all 81 configurations for ResNet-18 on an Ascend 310B1. For the 72 non-L9 configurations, predicted guidance scores correlated strongly with unified measured scores (Spearman ρ=0.831, 95% bootstrap CI [0.716, 0.898], p<0.001). The method matched the full-search recommendation at B=12. At B=18, it evaluated 22.22% of the space, reduced device-side evaluations by 77.78%, and returned the same recommendation as exhaustive search. The method therefore directs limited hardware measurements toward promising regions of the deployment space.