Threadripper 2990WX + 5 (7) RTX 3070 : Questions

Dear community, I am making tons of tries to get best of my settings and stuff that I allready own.

Got some improvments with :
Running TVAI on Win 10 → Win 11 keep constantly crashing whatever I do even with poor settings
Running unconventionnal use of PCIE → Putting GPUs in PCIEx16 leads to tons of crash also
SMT mode Off → TVAI seem to not like this, it crashes quickly, went back for SMT On, stable
So …I went for 5 risers PCIEx1 like this :

  • CPU : OPB settings : PPT=450, TDC=270, EDC=330
  • DDR4 : 4X8Gb quad channel 3600 MHz 16-20-20-20-38
  • PCIE : 4x PCIE Gen 3, 1x PCIE Gen 2, all of them on 1 lane
  • Everything is actively cooled with Delta @ 100%, jet noise around but t° look all fines.
    image

Encoding settings

  • Rez : 3860x2160, no upscale, same output rez, same fps
  • Model 1 : NYX HQ 60-80-20-40-10-20
  • Model 2 : Auto Theia D/F 20-10-60
  • Output : H265-MKV-Main10-Dynamic-Copy

Trying to denoise ST from best BluRay I own (no one has ever been denoised to Verhoven wishes)

  1. Running with SVIo on : is it of any use ?

  2. I can’t run the benchark, it crashes at the 4th test or close to : Is that a prove my setting isn’t stable or what ?
    If I don’t care about benchmarking, I’ve got somehow stable result however (picture below)

  3. Why me GPUs are so much unused ? No one of them reach 100% load.
    I checked PCIE bandwidth in AIDA64, either one PCI Gen3 or Gen 2. I, at most reach 70% low to 10%, which sounds to not being the key, but any confirmation will be appreciated.

  4. Is it regular to get allmost 50% load on a 32/64 CPU ? (Moreover boosted with OPB ). My setting is use “All GPUs”.

  5. In order to save PCIE lanes for future extension (+2 RTX 3070 or 4090), I use 1x SATA SSD(OS), 1x SSD for source file. I will ouput “Export as” through LAN on a RAID5 server 20 Tb drive
    Is there a way to save “Preview” files also thourgh LAN ? I did set “Export folder” on LAN : works fine when exporting but even if i set “Workspaces folder” to LAN drive, it stucks on local SSD, is it a bug or I am wrong ? Should I specify subfolder ?

  6. My RAM indicates it is loaded at most at 60%. is there any purpose to raise it to 64 GB ? It looks like not seeing at least 10 GB are free.

Best regards

What happens if you only use one GPU? If it runs ok, try disabling all but two GPUs and see what happens. If it still works ok, re-enable GPUs one by one.

I have three Radeon GPUs in my workstation. TVAI runs fine with one or two, but not with all three.

I’m guessing that TVAI either has problems with more than two GPUs or with odd numbers of multiple GPUs.

In task manager, you’re not going to see much more utilization that that. At least, that looks normal to me. The tensor core utilization does not always get reported as full use in task manager. They could be maxed out, but there’s still plenty of other things like CUDA that will not be used at 100 percent. So it’s not going to report that the GPU is totally bogged down.

I don’t know if they have changed things, but that hardware should do okay at running several instances of TVAI on several different videos at the same time. In TVAI 4, it cannot really use multiple GPUs on one video—not in the way anyone expects it to.

1 Like

TVAI 4 with multiple GPUs won’t produce usable performance improvements on a single video. It still shouldn’t be “constantly crashing” with them, though.

1 Like

I use TVAI 5.1, with those Tensor’s
It’s now rockstable, estimated time for encoding the whole movie 72h :smiley:
Any resume future that I have not yet seen ?

Waiting for the famous release to “professional”

Right, in the long list I can use hig CUDA use, which sounds to be good.
If models are truely efficient, to my mind they still still some more learning.

If using one it’s OK but ETA isn’t good enough for ivniting me to force more effort, I will be dead before encoding end
The drawback to me is that upper GPU is too close to the enormous heatsink, there’s a point where heat is very hard to remove.

If you know that one GPU is ok, the next step is to test it with two, then three, etc. You can use the Windows Device Manager to disable all but two, test for stability, fps and load, then enable a third, repeat tests, enable fourth, repeat, etc., etc.

From what Topaz has posted, “Professional” isn’t going to change anything about the way TVAI performs on a single video; it’s intended to improve performance on batch queues.

This is a bit newbie responde without being impolite. I am sure of the relaibility of my stuff, the problem is how to mix all of this with an AI enhancing purpose.
I just ended my first try and the result is simply outstanding. Some insane scenes unviewable on a large screen are 24’ look.

We will see for the “professional” release.

I don’t think the problem is your hardware, but compatibility between TVAI’s All GPUs mode and your particular setup. As far as I can recall from these forums, you and I seem to be the only ones here who have tried to use the app in All GPUs mode with more than two GPUs. And we are both having issues with it.

Very interesting, gonna focus it then ! Thanks for explanations
Waiting now for my M.2 to USB 3.0 adapter card ( +2 GPUs )
Also bought a Highpoint PCIE x4 to 4x USB 3.0 ( 4 GPUs PCIE x16 slot )

I need throughput … :sweat_smile:
Giving a try also on 13x AMD Navi with a low CPU to see if CPU can be “discarded”. It should be allmost useless if models are acurated designed for GPUs.

You’re right ! When using Nyx + Artemis, 3 GPUs look fully sleeping in task mananager.
Got a look in clock and bandwidth use, theay are all 5 used. not fully but as it was using Nyx + Theia

Windows is clearly reporting shit …

When you set TVAI’s AI processing option to GPUs, it still uses the CPU for everything else. File and project management, settings, etc. With nearly 100% GPU utilization my CPUs (2 of them) run at about 25%.

I track my GPU and CPU utilization with the AMD PRO performance metrics dashboard.

This app. is at most 2 GPUs written correctly, end fo story