GPU acceleration

[Dave Osbun]

Have you looked at the GPU Acceleration page at Sony’s website for Vegas Pro 12? All the info is there. There recommended (ie- tested) cards can all be considered lower-priced cards except the nVidia Quadro cards:
https://www.sonycreativesoftware.com/vegaspro/gpuacceleration

If you pay $170-$300 for a video card I don’t think you’ll see any rendering improvement over the cheaper cards they have listed. I don’t even see NewEgg carrying the GTX 570 or Radeon HD 6870 any longer.

If you do this for a living, i’d get the Quadro 4000 and nothing else.

Dave

[Dave Haynie]

When Vegas 11 first introduced full GPU acceleration, I bought both an AMD Radeon HD6970 and a nVidia GTX570 based board — neither was the fastest available, but they were $300 each. And that was nearly two years ago.

The HD6970 matched or outperformed the GTX570 at every benchmark, including Sony’s own. So that’s what I went with, and I haven’t been disappointed. You can certainly get faster versions of either GPU-company’s chips these days. And I suspect at least some of the problems were based on nVidia’s OpenCL implementation (with nVidia, CUDA is also OpenCL, so Sony says “CUDA-enabled”, but they’re really using the OpenCL APIs, which were not as debugged back then).

Given that, some things are still, years later, CUDA-only. Nothing from Sony, but some 3rd party stuff. So it’s important to know is you have anything critical that would benefit from CUDA but not from OpenCL.

It’s also important to understand that GPU acceleration isn’t a panacea. For one, it takes some time to compile a GPU problem, and some time to communicate to/from the GPU. So while your CPU may usually hit 95-100% alone, you may only see 70-80% use with the GPU. That’s the time the CPU is waiting for the GPU to finish things. What this implies is that, as your CPU speed increases and/or your GPU speed decreases, the GPU stops helping. So folks with fast i7 systems tend to downplay the effect of the GPU… meanwhile, guys like me (six core AMD 1090T at 3.2GHz) are slow enough to really appreciate the GPU.

There are also things it does well, and things it doesn’t. There are a couple things to be aware of. The GPU isn’t used much in just plain video decoding as part of a render. So if you’re rendering one video layer to another format, don’t expect the GPU to help much. Unless that video format itself is accelerated… most of the AVC CODECs get a boost on their own from the GPU. Once you start layering things and adding effects, the GPU starts being pretty helpful. I’ve seen 2x-3x improvements on some dense animation renderings, not too different in nature than Sony’s benchmark project (available on their site).

In short, like most things, it’s just another tool.

Also, watch recommendations for very expensive “workstation” GPUs, like nVidia Quadro or AMD FirePro, just in general. Yes, these are “optimized” for workstation use. But that generally means 3D mechanical CAD, since that’s about the only common workstation activity that’s really limited by OpenGL performance. And OpenGL performance is precisely what they’re optimizing here. Much of this is in software, not hardware. As well, at any given generation, the workstation cards may be using older technology. For example, until this year’s “K” (for “Kepler”) series of Quadro cards (K2000, K4000, etc), nVidia’s Quadro series were shipping with the same “Fermi” cores for the last three years, the GPU core that originated in the GeForce GTX480 in 2010. The consumer cards, GTX5xx and GTX6xx, where using the Fermi cores in early 2012.

So if you’re really planning to spend $700+ for a GPU card, at least be sure it’s not going to render video much slower than a $350 gaming card. All of these cards are doing video acceleration (which isn’t generally usable for editing, only playback) and compute acceleration (eg, OpenCL/CUDA), and this is what concerns you for video work. Nothing in the “professional” or “workstation” series are specifically optimized for video. Yes, some plug-ins (anything “3D”, for certain) do use OpenGL, and your pro-class cards have faster and more stable OpenGL than gaming cards. They may also have ECC memory and a few other things that say “professional”. You can even find a few cards, like nVidia’s Tesla cards, which offer really high-end GPU computing, cost more than all the rest of your PC put together (most likely), and don’t actually do graphics. Also not an issue in video computing.

-Dave