Hey Fellow vMix Enthusiasts.

I had the following issue last week. Once I started streaming I noted a GPU overload (GPU mem ~71%) and the stream started losing frames (down to 25-27fps). I suspected the video lower thirds and stinger, because I know vMix recommends converting to image sequences. I was afraid the system was going to crash, so as each talent finished their segment, I deleted their lower third input. Each lower third was taking 3% of the GPU mem! Once I deleted them all I was back down to 47%, and if I remember correctly the Overload indicator stopped, and the framerate went back to normal.

This was a completely last minute call, so I didn't have time to do the normal amount of prep or testing.

So, I set up a test today and converted all the lower thirds to image sequences (I didn't with the stinger since it has audio). To my surprise the GPU mem usage is now 96%! And this is before receiving vMix Callers or Streaming to any platform.

So, I'm coming to you hoping to be shown the error of my ways. Very gun-shy about animated graphics now.


Stream: 1080p 29.97
vMix Calls: 8 Callers
Lower Thirds: 8 (7sec)
Stinger: 1 (4sec)
Prerecorded Content: 3 videos (90sec, 5m, 30m)

Machine Specs:

AMD Ryzen 9 3950X 16-Core Processor 3.5GHz w/ 64Gb RAM
Nvidia GeForce RTX 2070 (High Input Performance Mode is checked)

Any tips or options to try would be appreciated.

-G

I do have to ask why does each talent has it's own lower third ? hope not because of the difference in name or so

You'll likely save some GPU cache, but I can't really put a number on it. 20% GPU mem is pretty decent as is. I have a 46 input project open right now, and it is 23% GPU mem currently.

This is the classic struggle with graphics designers versus character generators: the concept creation exceeds the practical implementation. I bet that's fun trying to fix a typo in the middle of the show with everything burned in.

I have an X-touch Mini (in MC mode), and I use one of the rotary encoders for lower thirds. The button press performs the keying operation and the rotary function allows me to do next and previous preset. I also have NDI Studio Monitor running on another screen with a direct output of that source so I can actually see the text. I used to put the lower thirds in preview to check, but I would then make the idiot move of hitting cut and delivering a lower third over black to the world.

You can create a shortcut and use SelectTitlePreset to select title, but it is by index so you'll have to be cognizant of the order of your presets when you create it.


That's interesting that it's an MP4, since that filetype doesn't support alpha channels. At that point, you really don't have any reason to convert it to an MOV and convert straight to an image sequence. vMix Video Tools should be able to do the conversion with no problem.

A good craftsman never blames their tools.

I've watched various generations of Chyron, VizRT, etc. absolutely eat it when a graphics department gets too greedy. Graphics departments who are not overly broadcast centric, meaning they scream "it worked in Adobe!" and nothing more, tend to heavily favor form over function. All character generator systems have their limitations, including the ones I just mentioned that cost the price of a small condo.

If your lower thirds are completely burned in and there's no need to add text, it leads to the question of why is it in GT at all when you can import image sequences directly in vMix?

I used to have a 30 second loop as an image sequence, and that posed no issues. That was also on top of the two 5 second stingers as image sequences. Again, no issues. The only reason I stopped loading the loop as an image sequence was because I got too impatient with the preset loading time.

Hey Babbitt,
My intention here is not to blame vMix, sorry if that's the impression I'm giving. I've been using it for under a year, so I usually find operator error is what keeps me from accomplishing my goals, and honestly that's my favorite problem to discover, as it's usually the easiest to fix. Hoping this falls into that (overflowing) bucket.

Also, I agree that the process of turning movs into GT Titles is a lot. On the original show sparking this discussion, I did convert the movs and input them as image sequences, and the result was GPU Mem 96%. I’ll try it again and see if I encounter the same issues.

As far as getting too greedy goes, this doesn't "feel" ambitious, but maybe it's an expectation problem on my end. Open to that as well.

Going back to your comment re last minute changes to client provided graphics, yes it can be a problem, but it's their problem, as they have approved everything I use, as opposed to mine. While I want to be as flexible as possible for them, I also need to limit my responsibilities so I can focus on the other inevitable things that come up before a show goes live.

Thanks for the conversation, you have given me a lot to chew on in this thread.

No offense taken. I think it was another character generator operator that introduced me to that saying... maybe even for a similar reason.

I will say that I once attempted to load up 7 stinger transitions, which would have probably put me at the 45 second mark on image sequences. It worked fine, but I was a bit uncomfortable with how little GPU mem I had left. So I never did that again afterwards.

You can try another video codec that supports alpha channel, like SpeedHQ from Newtek, and see how much a hit it takes on GPU mem. I know the Quicktime stuff is pretty severe, so don't bother with those codecs. It's been a while since I've tried SpeedHQ with alpha channels, so I don't remember how it affected the GPU mem.

If you get favorable results, you can then try placing all the those files in a list input, which in general practice is encouraged with regular video as you eat up far less GPU mem that way.

With the GT title related frame drops in fullscreen, I haven't noticed them myself, but I don't run in fullscreen mode that often. I'm personally reluctant to use fullscreen for feeding a critical screen. I'd rather take a latency hit using the Decklink output versus risking Windows toying with the fullscreen output. No matter what I do, disabling Aero Peek, etc., I still find ways to minimize or make the fullscreen output momentarily disappear.

Way back when, I used to have some of my GT titles stutter occasionally. I tried "Lock GT Surfaces" (I swear it was called something else in 23) in the advanced performance settings, and that seemed to help. But later I turned it back off and the stuttering never came back. So it reality I don't think that did anything.

One thing that may be worth trying is enabling High Performance Output Mode, which is also in the same area of settings. It did make a difference for me with frame drops in Multicorder. It also made a difference in another system I set up, which was a ridiculously overpowered dual Xeon workstation with a 2080 Ti.