DaVinci Resolve users have emailed us about their enthusiasm for 'alternative' GPUs for the Mac Pro. We measured the effect of the GPU on unrendered playback over a year ago. We figured it was time for another round of testing, especially with even more muscular GPUs available for the Mac Pro that run under OS X.
Blackmagic Design's DaVinci Resolve 9
This is an amazing pro app on two levels. First, it adds speed and power to color grading of HD video. Watch the demo video on the BlackMagic Design site to see what I mean. Second, it uses the GPU to apply and playback effects in real time -- no rendering required. However, the more effect nodes created, the slower the playback.
Our test project featured 8 clips of a 32 second 1920x1080 10-bit YUV 4:2:2 SDI 24fps movie. We set the maximum playback framerate to 200 fps. We wanted to see how fast each GPU could playback the movie. The first graph shows the average playback speed with no color correction or effects.
GTX 580+ = NVIDIA "Classified" (overclocked) GTX 580 GPU (3 power feeds, 3G VRAM) **
GTX 580e = NVIDIA eVGA GeForce GTX 580 GPU (2 power feeds, 3G VRAM) **
GTX 570 = NVIDIA GeForce GTX 570 GPU (2 power feeds, 2.5G VRAM) **
GTX 680 = NVIDIA eVGA GeForce GTX 680 GPU (2 power feeds, 2G VRAM)
Q6000 = NVIDIA Quadro 6000 GPU (1 eight pin or 2 six pin power feeds, 4G VRAM) **
R5870 = AMD Radeon HD 5870 (2 power feeds, 1G VRAM)
Q4000 = NVIDIA Quadro 4000 GPU (1 power feed, 2G VRAM)
R6870 = AMD Radeon HD 6870 (2 power feeds, 1G VRAM)
GT650M = NVIDIA GeForce GT 650M GPU (built-in Retina MacBook Pro), 1G VRAM
R5770 = AMD Radeon HD 5770 (1 power feed, 1G VRAM)
GTX285 = NVIDIA GeForce GTX 285 GPU (2 power feeds, 1G VRAM)
Ten PCIe GPUs are installed in a 'mid 2010' Mac Pro 3.33GHz Hex-Core. The GeForce GT 650M is built into the 'mid 2012' Retina MacBook Pro 2.7GHz Quad-Core i7. Both Macs were running OS X Mountain Lion 10.8.1. (Follow-up testing with 10.8.2 showed no signficant change.)
**Both GeForce GTX 580s, the GeForce GTX 570, and Quadro 6000 were loaned to us by MacVidCards. The GTX 680 and Radeon HD 6870 were unflashed "Windows only" GPUs.
Some of the GPUs started off strong but faded significantly when adjustment or effect nodes are added. Though all GPUs included in the test exceeded the 24 FPS target rate, adding a fourth node likely would have taken half of them "under water."
In our test last year using DaVinci Resolve 7, only CUDA capable GPUs (NVIDIA) could be used. Now AMD GPUs are supported but, as you can see, the best models compatible with OS X can't compete with the best NVIDIA GPUs.
We included the best MacBook Pro with the GeForce GT 650M discrete GPU to illustrate that it isn't up to the job except for very limited adjustments using DaVinci Resolve. For serious work, you'll want a Mac Pro with one of the 'alternative' GPUs.
DaVinci Resolve supports "slave" GPUs. For example, you could have a GeForce GT120 as your display GPU (no power feed required) and the GeForce GTX 580 (two power feeds) as your slave GPU to render adjustments and effects in real time. Or you could slave two Quadro 4000s or 6000s (one power feed each) to the GT120.
However, you need the full version of DaVinci Resolve to slave more than one GPU. The Lite version (which we used in this test) only supports one slave render GPU (in addition to the display GPU). We did try slaving one GPU but got mixed results. Based on our results, we think one "strong" display and rendering GPU is as fast or faster doing unrendered playback than using it as a rendering slave to a "weak" display GPU.