Performance and Power Analytical Models for GPUs and Mobile Devices.

摘要

Performance and power modeling and estimation are important for current and future processors. It is required to understand current processor's performance-per-watt and be able to estimate performance-per-watt for future processor architectures for a given workload with the lowest error margin possible. To achieve maximum performance with the lowest power consumption possible, it is ideal for any benchmark to operate at the highest level possible for performance-per-watt on a given processor configuration. In this dissertation, we present performance and power analytical estimation models that can estimate with 10% error margin for a variety of workloads and processor configurations. The models were developed to cover different processor configurations and architectures for different workloads such as cloud computation workloads, high-performance computation, 3-D applications, and Mobile Internet Devices typical usage model workloads. The models take different micro-architecture input parameters such as core frequency, memory frequency, number of cores, CPI, number of instructions executed, and processor efficiency. The baseline for our estimation models was Amdahl's Law, which scales by changing one micro-architecture parameter at a time (i.e. core frequency). We then expanded the performance model to change more processor micro-architecture parameters simultaneously and analyze the impact on performance. The models presented in this dissertation are analytical models rather than simulation-based. Given the amount of power the display unit consumes on a mobile platforms which affect battery life particularly on Mobile Internet Devices, we present different backlight inverter guidelines to reduce power consumption for LCD (Liquid Crystal Display) based systems (i.e. Netbooks and laptops).;Many performance and power estimation models were developed over the last decade, all aimed to estimate performance and power within 10% error margin covering a few workloads on a specific processor architecture, and in many cases, not offering visibility on the power consumption tradeoff for increased performance (performance-per-watt analysis and estimation). Our models cover different processor architecture parameters and a variety of workloads. Besides performance, we estimate power and performance-per-watt since there is a significant tradeoff between performance and power.