Dual Hetero Quadro on Heaven

Literally the punchiest title I could come up with. I’ve been told that heterogeneous GPU setups are ridiculously slower than a single-GPU. This is largely an anecdotal shrug of “hey - a second GPU doesn’t really slow my computer down in any meaningful manner at all!” I’m sure that I did this all wrong and that the GPU is capable of being tweaked into a setting where this all becomes conclusively - the legwork for that isn’t interesting to me so I haven’t done it.

TL: DR; A second GPU which doesn’t match up or SLI won’t give you cooties.


My workstation came with a NVidia PNY Quadro K600 GPU, which was replaced with a K620. I decided to put the K600 back into my computer as a secondary card and run UniGine Heaven benchmark to see just how slow a third wheel makes it. All tests were carried out at full-screen-exclusive resolution but were otherwise using the Heaven Benchmark’s default settings for their namesake. Stereoscopy was disabled for the non-stereo setting, and I used 3d Vision as UniGine’s the stereo method. For stereo settings, the 3D OpenGL Stereo profile was used with Stereo= Enabled and Display Mode= Generic Active Stereo.


I really can’t see any reason that a consumer would use this setup - it just amused me.

Configuration Detail Setting Stereo Score Frames Per Second Minimum FPS Maximum FPS
K620 Extreme Yes 131 5.2 3.5 24.4
K620 Basic Yes 269 10.7 5.2 21.6
K620 Extreme No 281 11.2 7.3 22.5
K620 Basic No 592 23.5 15.0 42.2
K620 + K600 Extreme Yes 131 5.2 3.7 10.0
K620 + K600 Basic Yes 270 10.7 8.2 18.4
K620 + K600 Extreme No 280 11.1 7.2 22.0
K620 + K600 Basic No 591 23.5 14.8 40.5


The relation between the numbers suggests that; * the second GPU does kind of slow it down measurably * the second GPU doesn’t slow it down noticeably * the Dual-GPU is a teensy bit helpful with Stereo rendering, but not quite worth the cost * in a real game, I could tell the second GPU to do PhysX work and maybe see some improvement?

I may have left some junk running on the desktop during the single-GPU tests, so the difference could be skewed a bit. Regardless; I’m confident that adding the second GPU to this desktop hasn’t sucked up all my PCIe lanes or something silly.

Questions; * how would this work in an AMD/NVidia mix? (Red/Green) * would an x86 OS work any good/evil with x86 benchmark? * would something with PhysX really be that different good/evil?

Peter LaValle avatar
About Peter LaValle
Peter is currently a PhD student at the University of Nottingham. His day work involves applying functional programming to problems with artificial intelligence. Someday maybe he'll retire and be a graphics programmer or demoscene coder.
comments powered by Disqus