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

Graphics Processing Units (GPUs) are rapidly bringing the computing

power traditionally associated with massively parallel supercomputers

into the mainstream devices we use today. They have the power to

revolutionize computing by enabling orders of magnitude faster and

more efficient execution of many applications. Unfortunately, many

modern applications and users cannot take advantage of the computing

capability present in today's GPUs because today's GPUs are used as

secondary devices to the much less powerful CPUs. As a result, the massive

computing power of GPUs gets wasted and underutilized for a large

number of important applications.

This project aims to take a fresh and comprehensive look at GPU design

with the goal of enabling GPUs as first-class computing engines that

can benefit an overwhelming majority of real-world applications and

users. To this end, this project systematically investigates the

hardware/software design space of three new execution models, which

progressively turn a GPU into an independent, first-class compute

engine in a hybrid computing system: 1) an enhanced master-slave model

where the GPU is able to perform multiple-application execution, 2) a

new peer-to-peer model where the GPU is autonomous of the CPU, 3) a

hybrid model where GPUs and CPUs are integrated on the same die and

are equals from the applications' and system's viewpoint. The project

comprehensively develops software, hardware and software/hardware

cooperative scheduling, resource management, and system design

techniques for all three models.

If successful, this project can pave the way to making GPUs

first-class computing engines used in all aspects of our everyday

lives for a majority of applications. Doing so is not only expected to

lead to much higher degrees of energy efficiency and user productivity

but can also potentially enable new applications and devices that can

take advantage GPUs.