Boost CPU Turbo Yield Utilizing Explainable Artificial Intelligence

This paper introduces a novel explainable artificial intelligence (XAI) framework aimed at efficiently enhancing CPU turbo yield by optimizing wafer acceptance test (WAT) parameters. This framework utilizes the Extreme Gradient Boosting (XGBoost) algorithm in the yield prediction model and incorporates Shapley Additive Explanations (SHAP) to explain the model's decision-making process. By calculating SHAP value, the importance of WAT parameters in the turbo yield prediction model is determined, enabling the optimization of the manufacturing process to improve yield. A comparative analysis is conducted between our proposed approach and the traditional methodology of Pearson correlation coefficient. In the Pearson correlation analysis, saturation current of WAT parameters is identified as the primary factor affecting turbo yield. However, it is observed turbo yield begins to degrade at high saturation current levels due to failing power constraints. The limitations of the traditional method in considering non-linear data correlation and multi-dimensional interactions of WAT parameters impede yield enhancement. In contrast, our method is employed to address these challenges. In the importance analysis of our proposed method, metal sheet resistance is identified as the primary factor impacting CPU turbo yield. Notably, our method accurately identifies the 6<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> metal wire from the 8 layers of metal wire in this product design, demonstrating its ability to highlight parameters related to the manufacturing backend-of-line process, even in complex and diverse IC layout designs of advanced node products. To validate the effectiveness of our XAI framework, production lots with adjustments to 6<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> metal wire are intentionally processed. The results confirm a 15% enhancement in turbo yield with a +2σ adjustment in metal sheet resistance. Overall, this framework offers an effective and comprehensive method for enhancing CPU turbo yield in the production stage.