Dynamic Superpipelining: Shaping Microarchitecture for Variable Frequency

动态超级流水线：塑造变频微架构

• 批准号: 0207785
• 负责人: Eric Rotenberg
• 金额: $ 22.92万
• 依托单位: North Carolina State University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2002
• 资助国家: 美国
• 起止时间: 2002-07-01 至 2006-06-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0207785&HistoricalAwards=false
• 关键词: Dynamic; Superpipelining; Shaping; Microarchitecture; Variable

Microprocessor performance enhancements tend to be inefficient in terms of power, especially as new sources of performance become scarce. Yet, performance remains a key competitive factor in the microprocessor market and inefficiency is somewhat inevitable. Microprocessors that support multiple frequencies are strategic because they can manage inefficiency. Performance requirements of systems change over time. Microprocessors with variable frequency can be used to balance performance, power, and energy by adapting frequency to changing performance requirements. This project predicts that the need for variable frequency will have a profound impact on the way microprocessors are designed in the future. Current hardware support for variable frequency is predominantly circuit and technology oriented: dynamic voltage scaling accommodates higher or lower frequency by raising or lowering supply voltage, respectively. In the future, support for variable frequency should also be an integral part of microarchitecture design. To demonstrate this overall vision, this project puts forth a novel microarchitecture called dynamic superpipelining (DSP). DSP switches from low to high frequencies by dynamically doubling the number of pipeline stages. DSP combines the advantages of two very different pipeline design styles (shallow and deep pipelines) that are optimized for different frequency targets. A flexible pipeline delivers better performance across a wider range of frequencies, ultimately improving the energy efficiency of the microprocessor. To evaluate DSP, it will be used as a variable-frequency substrate for a speculative frequency reduction technique recently proposed by the PI.

微处理器性能增强在功率方面往往是低效的，特别是当新的性能来源变得稀缺时。 然而，性能仍然是微处理器市场的关键竞争因素，低效率在某种程度上是不可避免的。 支持多个频率的微处理器是战略性的，因为它们可以管理低效率。 系统的性能要求随时间而变化。 具有可变频率的微处理器可用于通过使频率适应变化的性能要求来平衡性能、功率和能量。 该项目预测，对变频的需求将对未来微处理器的设计方式产生深远的影响。 当前对变频的硬件支持主要是面向电路和技术的：动态电压缩放分别通过提高或降低电源电压来适应更高或更低的频率。 将来，对变频的支持也应该成为微架构设计的一个组成部分。 为了展示这一整体愿景，该项目提出了一种称为动态超流水线（DSP）的新型微体系结构。 DSP通过动态加倍流水线级数从低频切换到高频。 DSP结合了两种截然不同的流水线设计风格（浅流水线和深流水线）的优势，针对不同的频率目标进行了优化。 灵活的流水线在更宽的频率范围内提供更好的性能，最终提高微处理器的能效。 为了评估DSP，它将被用作PI最近提出的推测性频率降低技术的可变频率基板。