CSR: Medium: Collaborative Research: Enabling GPUs as First-Class Computing Engines

图形处理单元(GPU)正在迅速将传统上与大规模并行超级计算机相关的计算能力带入我们今天使用的主流设备中。它们有能力使许多应用程序的执行速度更快、效率更高，从而使计算发生革命性的变化。不幸的是，许多现代应用程序和用户无法利用今天的GPU中存在的计算能力，因为今天的GPU被用作功能弱得多的CPU的辅助设备。因此，GPU的海量计算能力在大量重要应用程序中被浪费和未得到充分利用。该项目旨在以全新和全面的方式审视GPU设计，目标是使GPU成为一流的计算引擎，从而使绝大多数现实世界的应用程序和用户受益。为此，本项目系统地研究了三种新的执行模型的硬件/软件设计空间，这三种模型逐步将GPU变成混合计算系统中独立的、一流的计算引擎：1)增强的主从模型，其中GPU能够执行多个应用程序；2)新的对等模型，其中GPU独立于CPU；3)混合模型，其中GPU和CPU集成在同一芯片上，并且从应用和系统的角度来看是平等的。该项目综合开发了三种模式的软件、硬件和软件/硬件协同调度、资源管理和系统设计技术。如果成功，该项目将为使GPU成为应用于日常生活各个方面的一流计算引擎铺平道路。这样做不仅有望带来更高程度的能源效率和用户生产力，而且还可能使新的应用程序和设备能够利用GPU。

Graphics Processing Units (GPUs) are rapidly bringing the computingpower traditionally associated with massively parallel supercomputersinto the mainstream devices we use today. They have the power torevolutionize computing by enabling orders of magnitude faster andmore efficient execution of many applications. Unfortunately, manymodern applications and users cannot take advantage of the computingcapability present in today's GPUs because today's GPUs are used assecondary devices to the much less powerful CPUs. As a result, the massivecomputing power of GPUs gets wasted and underutilized for a largenumber of important applications.This project aims to take a fresh and comprehensive look at GPU designwith the goal of enabling GPUs as first-class computing engines thatcan benefit an overwhelming majority of real-world applications andusers. To this end, this project systematically investigates thehardware/software design space of three new execution models, whichprogressively turn a GPU into an independent, first-class computeengine in a hybrid computing system: 1) an enhanced master-slave modelwhere the GPU is able to perform multiple-application execution, 2) anew peer-to-peer model where the GPU is autonomous of the CPU, 3) ahybrid model where GPUs and CPUs are integrated on the same die andare equals from the applications' and system's viewpoint. The projectcomprehensively develops software, hardware and software/hardwarecooperative scheduling, resource management, and system designtechniques for all three models.If successful, this project can pave the way to making GPUsfirst-class computing engines used in all aspects of our everydaylives for a majority of applications. Doing so is not only expected tolead to much higher degrees of energy efficiency and user productivitybut can also potentially enable new applications and devices that cantake advantage GPUs.