This is the description of multi-GPU processing in Video AI Pro from Topaz Labs | Video AI Pro™ | Studio-grade video enhancement. For creative professionals. :
“Split-render video across GPUs or render multiple videos simultaneously. Video AI intelligently subdivides videos into tiles that run simultaneously across multiple GPUs. Scale to 2, 3, or even 10 graphics cards.”
If I’m reading this correctly, Pro is not doing true GPU multiprocessing the way the regular version tried to do but seldom succeeded, but is splitting videos, processing each part with a separate GPU, then stitching the parts back together when they’re done. IOW, just automating what a lot of us are doing manually now.
Oh, and the animated diagram of a GPU with air flowing through it that’s just below this? The air is flowing through the GPU in the wrong direction.
2 Likes
I might be wrong, but because of the mention of “tiles”, I think they are splitting the video in a different way than most of us might. Think of a video file as a Battenberg cake. Most of us would slice the cake in the usual way - equivalent to splitting the video in a temporal way. However, I think TVAI Pro may be slicing the cake lengthwise - hence the reference to “tiles”, and then reassembled. Both methods have their pros and cons. Anyway, whatever method is used, I think it’s worth reminding all users that both methods can be done in the standard version by using the command line, perhaps as part of a script.
Thanks.
Andy
Kyle posted a message about a “transition offer” for users who have been using multi-GPU in the 5.2.3 release forum. I contacted the help email, and they’ve issued me a no-charge Pro upgrade for the remainder of my current license. I’ll have 13 months to play around with Pro and then I’ll be done, because if there’s no alternative to buying a full license, I have no use for any of the other Pro features besides multi-GPU.
So far, I’ve only done a few test runs with the 5.2.3 release with Pro license, and the results indicate that multi-GPU on a single video is working pretty much the same way that it always has, about a 30-40% speed increase in benchmark, but real exports may actually be slower than with a single GPU. But I have Radeon GPUs, so no tensor cores here. That may make a difference.
I have a problem in general with the idea of scripting a division in a video. A script isn’t going to be able to identify the optimum point to make a split (at a hard cut or a fade-to-black scene transition), and we’ve seen builds of Video AI lose or duplicate frames during interpolations or at mark in/mark out points.
When I split videos manually, I set the mark in/mark outs to produce clips that overlap so I can sync them and make a cut through both that will produce precise splices.
I was really just wanting to make the point that the “Pro” version isn’t necessary to make full use of multiple GPUs if the user is willing to use the command line. ChatGPT or similar will give an outline of the required commands given the right question.
For those that share your concerns regarding video segmentation and concatenation, that’s a good reason to use the cropping and stacking (mosaic / tiling) technique instead. And that’s possibly why Topaz has chosen that route, if that’s what they’re doing, for multiple GPUs.
Thanks.
Andy
I’m not seeing any indication that splitting or tiling of single videos is happening. The temp file created during export seems to be the same as the one I was seeing in builds prior to Pro (no temp files of partial or tiled video frames), and the exports are taking the same amount of time as before.
I still have not gotten round to testing the multiple videos option, so I don’t know what happens there.
Hi, I am running Video AI Pro on an Epyc motherboard with Dual GPU’s
MZ32-AR0 Rev3
128GB
Dual 3080 Geforce GPU’s
Anyone know how much PCI bandwidth I need for upscaling and enhancement? Is there going to be a difference using PCI4.0 x16 versus PCI-E 1x to PCI-E x16 via a riser?
On a separate topic, has anyone tried mixing and matching different GPU models? Was thinking about throwing in an old 3060ti in with the dual 3080’s. Thought that I read something about Topaz AI Pro only working down to the slowest GPU speed? Any feedback would be appreciated.
