A Novel AI Based Approach for Performance validation

Ever increasing demand for high-performance processors and shorter time to market establish the need for faster performance validation strategies. But accurately predicting a slow pipeline unit has become extremely difficult with the increase in complexity of superscalar core designs and absence of error signature in performance validation. The current process is completely manual, inefficient, and time-consuming as it results in analyzing duplicate traces pointing to the same bug. It can cause a serious impact on product quality due to some performance issue that is getting masked by incorrect bucketing. In this paper, we propose a unique approach that leverages unsupervised AI techniques for performance validation which has shown positive results for Intel Core IP 10nm and 7nm projects. To cite on an average, we observe 93% reduction in total number of failures, resulting in only 7% to be debugged. This solution is scalable and can also be reused by other pre and post silicon performance validation teams like Atom, GT, and SOC