GPU Computing Resource Allocation and Prediction Based on Machine Learning Algorithms

This article proposes the BiGRU algorithm optimized by Transformer, providing a more reliable technical path for the fine-grained allocation of GPU computing resources in deep learning models. In the study, ExtraTrees, CatBoost, XGBoost, and LightGBM were selected as comparison models. The results showed that the proposed Transformer BiGRU model performed the best in deep learning GPU resource regression prediction. Among them, the MSE of this model is 2.208, significantly lower than ExtraTrees' 4.414, CatBoost's 6.71, XGBoost's 4.464, and LightGBM's 2.942; The RMSE is 1.486, which is lower than the 2.101, 2.59, 2.113, and 1.715 of other models; MAE is at its lowest level at 0.969, which is better than XGBoost's 1.442 and LightGBM's 1.296; MAPE is 26.938, which is lower than ExtraTrees' 34.478, CatBoost's 37.379, and XGBoost's 36.679; R reaches 0.855, higher than LightGBM's 0.803 and XGBoost's 0.69. This indicates that the model has smaller prediction errors and better fitting effects on GPU resources. Its research value lies in laying an important foundation for improving the accuracy and scientificity of GPU computing resource allocation, and has a positive significance in promoting the efficient utilization of resources in the field of deep learning.