I suspect eventually Topaz will abandon V3 as we know it and go back to an improved V2 instead.
Rebooting is useless, now I took a 4mins clip for the episode, the same settings as before:
This software is a meme.
These are the settingsā¦
I use Apollo because itās a fight scene with fast movements.
And I add some grain to prevent the image from losing its natural touch.
The duration could be correct because upscaling and frame interpolation will be 2 separate operations and 4k@60fps are a lot of pixels to process (x10 times more than the original 1080p@24fps).
Also, the CPU + RAM plays a big role when upscaling not just the GPU.
To check whether there is anything unusual with your PCs performance u could post the result of the TVAI built-in-benchmark here.
The upscaling to 4k + changing to 60fps will be bottleneck by your CPU + RAM (should be above 32 in this case) and your Hard Drive (faster will be better since TVAI breakdown the work in small files, in a processa that will be more or less like: write file (more than one at time) ā fill (work) ā combine ā delete ā next file (piece)). This happens in your temp folder inside the %appdata% (for windows users) and this part will be the starting point to bottle neck everything.
If the process you doing require more RAM that your system have, it will also use your OS disk to swap the extra memory that was required. In my tests, 32 ram + SSD, for most cases, is āokā (a.k.a your system will not freeze completelly) to work with 4k 30, not sure 4k 60.
So even if you have a monster GPU, the way they design the workflow will always be 100% bound with: read and write speeds (and bandwith) of your SDD/HDD, your CPU handling it and the ammount of RAM. That is why you probably feel that your GPU has almost no usage during the process.
Iāve already posted it, a lot of timesā¦ but well, hereās again:
4k Benchmark
Topaz Video AI v3.2.8
System Information
OS: Windows v11.2009
CPU: 13th Gen Intel(R) Core(TM) i9-13900K 31.685 GB
GPU: NVIDIA GeForce RTX 4090 22.096 GB
GPU: Intel(R) UHD Graphics 770 0.125 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 3840x2160
Benchmark Results
Artemis 1X: 06.47 fps 2X: 03.45 fps 4X: 00.48 fps
Proteus 1X: 06.10 fps 2X: 02.38 fps 4X: 00.52 fps
Gaia 1X: 02.91 fps 2X: 02.03 fps 4X: 00.96 fps
4X Slowmo Apollo: 11.60 fps APFast: 21.31 fps Chronos: 06.38 fps CHFast: 10.37 fps
What are you talking about? This is the performance monitor while doing the same process again.
Thatās why I made sure to set the temp folder in the same place as the output video location, in case that āwill make a bottleneckā.
Interesting. By your performance monitor, the bottle neck probably happens in your CPU. It also seems that your system has a very good balance since you can output 78% of your GPU without things piling up in your RAM and SSD. Iām not sure what you could change to squeeze more performance.
In a similar scenario, my RAM tends to stay almost full, and my CPU around 70% and GPU around 30% (I definitely need to upgrade my CPU. It is weird that my 5700G can handle 4k in premiere without GPU, however when I need to export/render even with 3090 cuda cores and faster RAM, it takes longer than the FX8320 (without cuda). So probably there some instructions or else missing in my CPU).
2 Likes
The problem is obviously with the programme, youāve seen my PC, it has plenty of power. Itās something Iāve discussed many times over the last few months here with several people, most of them prefer to cope with it and not complain. I use more AI programs in relation to video post-processing, either for interpolation or upscaling. And they are all much faster and do a much better job with the hardware.
The high CPU consumption doesnāt make much sense either as I have chosen NVENC for encoding and it should be the graphics card that encodes the video. That 78% GPU usage you see doesnāt make much sense either, itās actually a pretty low usage when the usage should be maximum and constant, instead of having so many peaks and troughs.
CAIN, in its TensorRT version, gives me around 200fps in a 2x interpolation to a 1080p videoā¦ do you know how many FPS Topaz Video AI gives me with the Apollo model which is basically like CAIN (at least by the looks of the interpolation artifacts in the output video they are exactly the same) on the same video? 25 measly FPS.
In video upscaling itās the same, only Topaz Video AI has the best video upscaling for IRL/CGI at the moment, thatās why I have to use it. But without going any further, when scaling an anime opening with Topaz Video AI itās very slow, while using RealCUGAN itās much faster.
I have hardly any idea about AMD hardware, but I can tell you that itās probably a TVAI thing, and that if it were better optimised it would work much better for you.
PS: Oh, and I might add, with those other AIs I use, the CPU usage is practically zero (unless I choose CPU encoding), the GPU does all the work, which TVAI doesnāt as you have seen.
2 Likes
I noticed when comparing your benchmark results, that my 7950X + 4090 scores for Apollo and scaled Proteus on 4K differed much more than what would be expected due to the similar system. Maybe there is something to optimize on your system in the range of 30-50% for your task.
Topaz Video AI v3.2.8
System Information
OS: Windows v10.2009
CPU: AMD Ryzen 9 7950X 16-Core Processor 31.118 GB
GPU: NVIDIA GeForce RTX 4090 23.59 GB
GPU: AMD Radeon(TM) Graphics 0.47445 GB
Processing Settings
device: 0 vram: 1 instances: 0
Input Resolution: 3840x2160
Benchmark Results
Artemis 1X: 08.89 fps 2X: 03.76 fps 4X: 00.69 fps
Proteus 1X: 06.92 fps .2X: 03.60 fps 4X: 00.95 fps
Gaia 1X: 03.34 fps 2X: 02.32 fps 4X: 00.95 fps
4X Slowmo Apollo: 17.58 fps APFast: 19.26 fps Chronos: 07.28 fps CHFast: 13.65 fps
2 Likes
Do you have above 4G decoding and Resizable BAR enabled in the Bios ? If not enable those 2 and see if you get more juice from TVAI
Also try to switching hardware scheduling on/off
Thx for the bench.
8 fps is like meh!
When i start doing video stuff about my photography i think i will go FHD instead of 4K because of the performance.
I cannot believe is almost 6 months and the hard crashes on the Mac Pro 2019 still continuesā¦
No, its not macOS, 2.6.4 doesnāt crash running all night with the same temps, vram, etc.
None of the other video software i use crashes the entire system like VAI
That was already enabled by default on my motherboard:
And the GPU scheduling was enabled before, now Iāve disabled it, restarted Windows andā¦
Benchmark @ 4kā¦
Topaz Video AI v3.2.8
System Information
OS: Windows v11.2009
CPU: 13th Gen Intel(R) Core(TM) i9-13900K 31.685 GB
GPU: NVIDIA GeForce RTX 4090 22.096 GB
GPU: Intel(R) UHD Graphics 770 0.125 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 3840x2160
Benchmark Results
Artemis 1X: 08.97 fps 2X: 03.81 fps 4X: 00.60 fps
Proteus 1X: 06.97 fps 2X: 03.34 fps 4X: 00.66 fps
Gaia 1X: 03.06 fps 2X: 02.11 fps 4X: 01.05 fps
4X Slowmo Apollo: 15.85 fps APFast: 20.87 fps Chronos: 07.01 fps CHFast: 12.21 fps
Same bsā¦
There is nothing I can do, and there is nothing that can be done in TVAI because the only thing that is offered is the possibility to change the maximum GPU consumption (at least if the streams could be increasedā¦ maybe that would fix the problem.), which I have always set to maximum and it never uses the maximum. I have already mentioned it several times, the problem is at the internal level of the program, because it uses very little VRAM memory, and it should use much more.
Iāve loaded up a video and same settings as you and Iām getting 4.2fps on a 3070
See? This program canāt take 100 advantage of a 4090.
I remember years ago when I had a 1080Ti, it performed much better than a 2080. 
1 Like
Your 2400 DDR5 is slow. Why buy a top of the line CPU and GPU only to put the slowest RAM in it?
Maybe with the money you havenāt spent on buying a license and supporting the developers you could buy some decent RAM.
Are you 100% sure you have selected the 4090 under preferences ?
My RAM is 4800Mhz, and DDR5 slowā¦ yeahā¦
What are you talking about?
Obviously, and Intel QHD Integrated graphics with 0.125 GB of VRAM couldnāt even handle 480p video.
I meant 2400 running at 4800 MHz. Itās pretty much the slowest spec DDR5 there is.
You must have at least $300 available to buy some 5600 or 6000 MHz RAM since you havenāt bought a Video AI license.
Yeah, definitely youāre a troll. I wonāt display any message from you anymore.
1 Like
