Artificial neural networks for voltage-frequency prediction using on-die measurements

• Main Author: Kannan, Ponmithiran
• Other Authors: Mohamed M. Sabry Aly
• Format: Thesis-Master by Research
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/169942
• Institution: Nanyang Technological University

Every APU, GPU and CPU device produced by AMD operates across multiple voltage regions known as P-states. The product specification determines the frequency of the clocks at each P-state. To ensure that the device is supplied with sufficient voltage to operate at the given frequency, we build part-specific voltage to frequency curves through a methodology known as adaptive voltage frequency scaling (AVFS). AVFS can characterize the device’s voltage frequency relationship using on-die critical path oscillators (CPOs) that provide parametric frequency measurements of the internal circuit paths. Depending on the die size, products may have 100s to 1000s of CPO measurements available for each device. The current methodology predicts the frequency at a given voltage using a solver-based linear regression where CPO measurements are the features and the characterized frequencies (Fmax) are the labels. Despite the efficiency of the supervised linear regression, the prediction errors still have significant room for improvement. The reason is that linear regression does not adequately address the non-linear relationships between CPOs and Fmax hence introducing inductive bias into the prediction model. Moreover, critical non-parametric information such as core identifiers and die location are omitted from the algorithm. This report aims to capture the extensive research on performance prediction over the past year and summarizes the progress on breakthrough deep learning algorithms that substantially reduce the prediction errors across all voltages, thus allowing AMD to squeeze out even more performance than previously thought possible. We have identified the major gaps in prediction algorithms through an extensive literature review of the semiconductor industry and internal intellectual property. One is the exclusion of categorical (non-numeric) information and the other is the overdependence on linear regression for the prediction logic. Hence we evaluate the application of deep learning approaches on conventional prediction tasks to highlight the apparent benefits of recent advancements in machine learning.