I’ve noticed that newer versions seem to rely almost exclusively on CPU, rather than GPU. Under heavy workload, my CPU pegs out at 100% while my Nvidia RTX 4070 shows almost no usage at all. AI processor is set to Nvidia and I’ve checked with several different monitoring apps, and they all show my GPU at close to idle. The screenshot below was taken running 2 processes simultaneously, upscaling 576p to 1080p. What’s happening here? Is CPU the bottleneck now?
The GPU usage graph in Task Manager is highly unreliable.
Please use GPU-z to check the GPU usage.
From your screenshot, the GPU temperature shown in Task Manager is 55°C, which isn’t normal for a 12% load. I suspect the actual GPU load is higher than what Task Manager is indicating.
Also if your CPU load is at 100%, you better set simultaneous process to 1.
If you are sure the TVAI is not using the GPU as expected, you may try switch the AI processor to Auto in Perference. Sometimes it will fix the problem.
If this doesn’t fix the problem, you may want to uninstall the TVAI, and reinstall again.
I was running 2 processes for demonstration purposes only. That’s not my usual workflow.
I also tried CPUz as one of the many tools I used to check GPU usage, so I’m pretty sure 12% is reasonably accurate. This has been shifting gradually over the last few releases and finally just got to a point where it’s all CPU. Today is the first time I needed to upscale anything in a couple of weeks and I was struck by how the shift to CPU is complete now.
Did not realize there was an Auto now, so I’ll try that, but I would be surprised if it changes anything.
Interesting. What version of TVAI are you running, what input and output resolutions, and what model is used?
this is my hw utilization running two parallel 2 x artemis medium upscales at once for two 460x360 clips.
You have a pretty good CPU, but it could well be that you’re CPU bottlenecked.
As you can see from my utilization, myself I’m bottlenecked on something , so my GPU and CPU is only being used about 30% each. It doesn’t seem to be on the hardware level though, since RAM bandwidth is only 50% utilized (2 channels @ DDR5 6000 = 96 GB/s transfer speed). As such it must be the VAI code itself
As you can also see, 8 cores are being used at average for these two parallel inference processes, as reported by the detailed hwinfo metric, so in my case the CPU isn’t peaking. However that is for this specific VAI model being used, and the specific input resolutions. I’ve noticed for other models or resolution, sometimes it uses 13-14 cores, so the specific input resolution and model seems to matter a lot. If you could share the details I asked for initially, I could see if I can replicate your findings. If I can, then we know that indeed, that specific configuration is a CPU hog, and there’d be nothing we’d be able to do about it.
Also, as Jacky pointed out, GPUz gives you a bit more accurate information. If you’re only relying on Task manager, then at least make sure to switch the GPU view to “Cuda”, since that’s what VAI uses to do the actual frame crunching. IIRC the default task manager view is “3D”, and the overall GPU average to the left is rather useless. Only Cuda usage matters in terms of GPU utilization for the VAI workload.
So I misread Jacky’s initial response and thought CPUz and not GPUz, not realizing it’s a different tool. The results are slightly better at around 35% GPU load with 2 processes running, but still not the distribution I’m used to seeing. In older versions, I saw more or less the reverse - 100% GPU and 35% CPU.
This test is running 2 instances of Proteus Auto, upscaling 720x576 by a factor of 2x to 1440x1152, output to ProRes HQ. No other actions. Version is the newest @4.0.3.
I did have an issue with the 4.0 release or soon after where it was hammering my cpu and not my GPU. A subsequent update fixed it. I am also on 4.0.3
I only slightly perused the thread - have you tried running only a single task at a time? Running multiple tasks at the same time hasn’t really worked out for me.
