Bridging the Gap Between CPUs and GPU with Out-of-Order SIMT

Most of today’s computers, from cell phones to supercomputers, are heterogeneous: they integrate processors that are optimized to quickly execute a few tasks (CPU
cores), and processors that can perform many independent tasks in parallel (GPU cores). GPU cores and CPU cores have different instruction sets: they understand
different languages. A task written for CPUs cannot run on GPUs, and vice versa. As a result, programming current heterogeneous architectures is challenging and few
applications can take advantage of the processing power offered by GPU cores. We will address this incompatibility by designing a hybrid CPU and GPU core, which presents the same instruction set as multi-core CPUs while offering the same parallel performance as GPUs.
To experiment and validate our proposition, we will model this new hardware in a software simulator mimicking the behaviour and timings of the proposed hybrid
architecture. We will develop the design using an existing CPU simulator.
This project will significantly ease the development effort required for designing applications which run efficiently on modern computers. It will also harness the
underutilized processing power available in contemporary hardware.

Faculty Supervisor:

Kaamran Raahemifar

Student:

Anita Tino

Partner:

Discipline:

Engineering - computer / electrical

Sector:

University:

Ryerson University

Program:

Globalink