Introduction to the GPGPU Benchmarks
Now that OpenCL is established as a multi-platform GPU-computing API, it's relatively simple to benchmark general-purpose GPU application performance, as long as the applications are truly platform agnostic. Previously, when the GPGPU universe was divided into CUDA (Nvidia) and Stream (AMD), we faced the problem that most applications supported only one of the two environments, and could thus not be directly compared to each other.
Apart from synthetic benchmarks, Bitcoin mining is one of the few practically useful, albeit somewhat narrowly focused, examples for GPU computing. A problem of using Bitcoin mining as a benchmark is that the existing OpenCL miner will typically not support a new graphics card generation upon release. Thus, we have to keep updating the OpenCL miner and benchmark results.
LuxMark is the second practical GPGPU application in our benchmark suite. It's based on the free render program LuxRender. A second version of this benchmark is now available, but we still use the well-known scene LuxBall HDR; we want to test a broad range of graphics cards, and this scene is the least-complex one. The other two test scenes are too difficult for entry-level cards. Since we want you to be able to compare the GPGPU scores of graphics cards with the GPGPU scores achieved by APUs and IGPs, we don’t plan to replace this popular scene with another one until after 2012 at the earliest.
GPU Caps Viewer
GPU Caps Viewer is a synthetic benchmark with an interesting mix of computation done via OpenCL and post-processing and graphical output with or without anti-aliasing. The PostFX test is a variant of the Nvidia demo for the API call oclPostprocessGL in the company's GPU Computing SDK, where a blur effect is added in a post-processing step after the image has been rendered.
The particle test is traditionally one of Nvidia's strengths. However, AMD’s new GCN architecture is catching up.
The so-called N Queens Problem has nothing to do with Freddy Mercury; it's a complex mathematical chess problem. Eight queens need to be positioned on a chess board in such a way that none of them could take another one according to chess rules (but disregarding black and white). The goal is to obtain the full solution set in the minimum time.