Application Benchmarks: ParaView & VisIt

SpecViewPerf is a good start, but I felt the need to try some more ‘application’ benchmarks.  To do this, I worked with the developers of two popular open-source visualization packages: ParaView & VisIt.  I obtained a copy of the Supernova Dataset used in the recent Ultrascale Visualization paper, and extracted one default isosurface (in the median of the data) and constructed a simple flyaround.  I disabled all attempts at level-of-detail, and forced displaylist generation.  This way, the geometry is loaded to the video card and the problem becomes a pure geometry-pushing problem: How fast can it push triangles through the pipeline to the screen?

I used FRAPS to track the framerate over the runs, and wound up with results like this:

You can see easily how the Quadro5000 doubles the performance of the previous generation 4800 and 5600.  One interesting thing to note here is how the 5600 and 4800 benchmark almost identically.  During the early discussions about the 5000 and the Fermi architecture in general, NVidia gave us some background about the last few generations of hardware.  With the sudden success of GPGPU and CUDA, and the advent of larger screens, much of their energy has been devoted to increasing the ability of the hardware to simply push pixels (Fill Rate).  While this was great for many applications, raw geometry performance stagnated.

The Fermi architecture, combined with Quadro, brings a 5x increase in raw geometry performance.  This is even easier to see in some Maya Benchmarks.

PreviousPage 3 / 6Next
Tags