Surrogate-based Analysis and Design Optimization of Power Delivery Networks for Cloud Computing Server Applications
Cargando...
Fecha
2020-09
Autores
Leal-Romo, Felipe
Título de la revista
ISSN de la revista
Título del volumen
Editor
ITESO
Resumen
Descripción
As cloud computing escalates its usage model, it is very critical for vendors and companies that provide this service to reduce the total cost of ownership (TCO). By looking thoroughly into cost reduction opportunities, one of the most promising areas of opportunity is computer server’s power consumption. This means, the design of a robust hardware capable to perform adequately during high computing demand, avoiding large computer’s burning power translated as heat dissipation, and in consequence, more power consumed to keep datacenters cool. By taking this precedent as a baseline, power delivery (PD) engineers’ job is taking more relevance to provide reasonable recommendations that do not increase bill of materials (BOM) cost of motherboards and packages, and maximize performance avoiding excessive power consumption. In this doctoral dissertation, PD analysis is presented as a discipline that implies trade off decisions between performance and cost savings. It is demonstrated how the implementation of design of experiments (DoE) and space mapping (SM) can be very effective numerical approaches to aid PD engineers taking quicker and informed decisions, while optimizing the design of power delivery networks (PDN) for computer server’s design. To achieve a systematic optimization of the PDN, this PhD dissertation proposes a general methodology to implement surrogate-based models. First, by exploiting parameters extraction (PE) for lumped circuit PDN model to fit impedance profiles, and second by generating black box surrogate-based models to enable fast and accurate optimization of the PDN performance. For black box surrogate-based models’ generation, this dissertation compares four different machine learning techniques: Kriging, generalized regression neural networks (GRNN), polynomial surrogate models (PSM), and support vector machines (SVM). The best resultant metamodels are exploited for fast PDN optimization in two different industrial applications: a voltage regulator (VR) implementing dual power rail remote sensing, intended for communications and storage applications, to find the optimal sense resistors and loading conditions; and a multiphase VR of a server motherboard, by finding optimal compensation settings to decrease the number of bulk capacitors preventing CPU’s performance loss
Palabras clave
power delivery networks, optimization methods, surrogate-based techniques, cloud computing servers, computer design, voltage regulator, impedance profile, voltage drop, design of experiments, Kriging, Support Vector Machine, Parameters Extraction, Design Optimization, Polynomial Surrogate Models
Citación
Leal-Romo, F. (2020). Surrogate-based Analysis and Design Optimization of Power Delivery Networks for Cloud Computing Server Applications. Tesis de doctorado, Doctorado en Ciencias de la Ingeniería. Tlaquepaque, Jalisco: ITESO.