Empirical Evaluation of Fused CPU+GPU Architectures

From Master Projects
Jump to: navigation, search


About Empirical Evaluation of Fused CPU+GPU Architectures

  • This project has not yet been fulfilled.
  • This project fits in the following Bachelor programs: {{#arraymap:|, |xXx|bachelorproject within::xXx|,}}
  • This project fits in the following masterareas: {{#arraymap:Parallel and Distributed Computer Systems, High Performance Distributed Computing|, |xXx|project within::xXx|,}}


Description

Traditional GPU-enabled systems suffer from an inherent bottleneck in the communication between CPU and GPU (usually done by PCI-express) due to the physically separated memory spaces. Fused CPU+GPU architectures aim to solve this problem by encapsulating the GPU on the same die as the CPU, and allow the two to share the same memory space. However, in such a design, the size, the capability, and the memory available for the GPU are more restricted than those specific to discrete GPUs. In this project, we aim to investigate the performance of such platforms - currently produced by both AMD and Intel, and featuring in many laptop, desktop, and server solutions. Specifically, we are interested in answering the following questions: (1) how do these architectures perform for traditional GPU benchmarks (see Rodinia, Parsec, Lonestar), (2) what are the true limitations of the GPU platforms in such fused architectures (by means of microbenchmarking), and (3) do these platforms really solve the CPU-GPU communication bottleneck? This project will require a moderate (low-level - C, OpenCL) programming load and a high experimental load (i.e., lots of experiments to be designed, conducted, and evaluated) .