I read that you can get better performance by running multiple instances of Video Enhance AI.
I also read that the process takes no more than 6GB VRAM (at least for an SD/720 to 1080P upscale - presumably it uses a lot more the higher up you go).
Question for someone that actually has an RTX 3090 or A5000 (something with 24GB VRAM). At 6GB a pop, you should be able to run 4 x instances of VEAI. Is this correct and is it roughly 4 x the performance of one instance, provided you have a decent CPU (I’ve got a 10850K, technically I could assign 2 x 4.8Ghz cores to each instance and have 2 left over for general system usage).
We use a Ryzen 9 5900X 12-core; 128gb RAM and an RTX3090.
I did the test this morning: I can’t get better performance with multiple instances of the software. With two instances, it is twice slower, with three instances, 3 times slower.
For info we’re porcessing Sony Venice-made 6K files (converted to Prores422HQ on input).
Maybe I’m doing something wrong, though. If someone has a tip, I would be more than happy to learn…
Well I did another test on another system, and it is working well: the other system is a 24-cores/48threads Threadripper with 128Gb RAM and TWO graphic cards: one Quadro RTX6000 and one Quadro RTX4000.
Now with “Use all GPU” enabled, I can launch 2 or 3 differents instances of VEAI and I get a huge boost in performance → you only get better performance if you’re using different instances if you’re using more than 1 GPU.
So the key is to have at least 2 GPU: with that combo of an RTX6000 + an RTX4000 (which are, under the hood, really just a Geforce Titan (ie: 2080TI with 24Gb RAM) and a Geforce 2070 = ok cards but not the absolute top-notch), I get almost TWICE the performance of a single 3090, it’s crazy!
So I guess that with 3 budget-friendly GPUs or more, one would really get insane performance.
I found that launching 3 instances of VEAI gives you the best performance. It’s a little better than two instances, and much, much better than only one. 4 instances gives no advantages and is less stable (uses more CPU).
Note that you can leaves all instances with the “use all GPU” option. I tried asking one instance to use the RTX4000 and the other the RTX6000: it’s not good, performance is poor.
And just for fun I also tried the “CPU” mode on the Threadripper: well, uh, don’t: it is at least 30x slower than on the GPU.
I think we’re now more or less building a “render farm” of VEAI. In the following months, we will probably build a machine with as much GPU as we can cram inside, so we can speed up the process a bit. Will update!
