I’m currently using the trial version of TOPAZ VIDEO 4.0.4 and I’m curious about the hardware upgrades, specifically CPU or GPU, that would yield the most effective and efficient results for video upscaling/encoding. After some research using the search options, it appears that users with the latest and most powerful hardware, such as the NVIDIA RTX 4090 GPU, may not see significant benefits from the software. This seems to be because TOPAZ Video AI (TVAI) doesn’t fully leverage or maximize the potential of high-end GPUs. Upgrading to the latest hardware may not deliver optimal results unless the software undergoes changes to effectively utilize the additional GPU processing power.
Hence, my questions to the community are:
What GPU, CPU, and RAM specifications do you recommend for achieving the best performance with TVAI?
Are there any other AI programs that can fully harness the capabilities of current hardware, providing faster and superior upscaling?
For question number 2: I don’t explore much of what other AI upscalers can do. I have seen other people on there forums report that TVAI is the best they have seen. Last report like that was a few months ago. It might change when big companies start offering AI upscalers.
CPU : 14900k/13900k
GPU: RTX 4090
RAM: DDR5 (as high as it support)
Cooling is important as well, you may want 360 AIO for CPU cooler.
Not that I am aware of.
There are other AI programs for Video enhance but last time I tried they were not as good as TVAI, but that was long time ago, you may try it yourself to see which one you prefer.
Depends a bit on the processing options. Upscaling a video to full HD right now is only moderate power hungry (RTX4090). The CPU (13900K) eats 125 W in the long run when set up according to Intels specifications.
I think they will preform the same. I found one 8GB user benchmark a few versions ago. Comparing against the other user that posts their 16GB every little update, they look the same.
it appears that users with the latest and most powerful hardware, such as the NVIDIA RTX 4090 GPU, may not see significant benefits from the software
There seems to be a threshold beyond which having very advanced hardware will have little impact on the performance of this application, although I’m not sure what that threshold is. Other users have commented on this as well.
Are there any other AI programs that can fully harness the capabilities of current hardware, providing faster and superior upscaling?
This is relative to what you are trying to accomplish. One area where TVAI has thoroughly struggled for me is with animation. But, I use another AI application that processes animation perfectly. Other areas where TVAI has been sub-par for me are with deinterlacing and denoising. But, again, I use different apps for these things.
Get the fastest RAM you can. I had DDR5 4200 and was bottlenecked on it using a Ryzen 7950X and a GTX 4090. Managed to get it up to DDR 6000 now, but am still bottlenecked on the memory controller. Especially when running multiple jobs simultaneously. Two jobs and they saturate the memory bandwidth, leaving the GPU run only at about 30% utilization and CPU at around 40-50%
Get the fastest CPU you can instead of the fastest graphics card if you’ve already bought fast memory to alleviate that initial bottleneck. The best GPU on the market is overkill compared to the best CPU available as my utilization numbers above indicate. TVAI does a LOT of work on the CPU.
Not much VRAM (video memory) seems required. I mostly upscale SD footage to HD, and don’t work on 4K or 8K material. If you do the latter, then more GPU memory may be beneficial, but if not, then a 16GB VRAM card should be more than enough for at least 2 simultaneous jobs.
Yes. Plenty.
Pretty much all my other ML workloads use 100% GPU all the time. Any model I run with ONNX, TensorRT or just pytorch uses all GPU resources. As for programs, they are in-house developed mostly, but if you’re looking at for example AI image generation like “stable diffusion”, then that’s a use case that will use all your GPU capacity no matter how much you throw at it. It really depends on what sort of problems your need to solve / use cases you have.
As for the future of Topaz products, who knows. it could well be that Topaz eventually re-architects some parts of their software stack so that they don’t need to shuffle so much data back and forth between CPU and GPU all the time. Their current architecture is very modular and I understand why they chose that; it’s good in a lot of ways, but not when it comes to efficiency and throughput. There’s no technical reason why they simply couldn’t short-circuit some operations directly on the GPU, thus increasing the GPU usage massively, but their engineers would have to speak to the reasons why they haven’t (yet). My guess; feature priority and maintenance cost rather than performance as priorities.
For newer content or content that does not have a lot of fine details, I have had a lot of success with Real-CUGAN.
For older animation, or cartoons that have many fine details, I have had success with VideoProc/Winxvideo (same program). Their Reality model is not very good for Reality, but it ironically works fairly well with animation, as does the animation model (which may very well be based on Real-CUGAN).
Otherwise, I have had the most success with Iris - although Iris LQ tends to remove some of the saturation and hue variation from the cartoons.
