no worries, we are not deleting your post. But I would like to move it the the Virtual GPU category if that is ok. Because I am certain there you will have a much better chance of getting the attention of experts.
Hello, thank you for the advice. I switched from best effort to fixed share because dividing the card into 4 parts of 12-Q (12GB) in best effort mode really makes everything good, but if 3 people start heavy games (first profile - Cyberpunk 2077, second profile RDR2, third FarCry 6, fourth Dota 2), then everything immediately drops to 28-30 FPS. If there is a fixed share, it guarantees that at least some games will run at 60+ FPS, except for Cyberpunk and FarCry RDR2, they immediately drop to 30FPS :(
You also said to delete FRL, after rebooting the virtual machine, this flag is set again and the machine works again at 60fps… I don’t know what can be done here to completely remove such a thing.
I don’t have experience in terms of which is best to use. Do you know how to better adapt the scheduler for my tasks?
I came across an old T4 card yesterday, which was set as passthrough. It has 16GB and 2560 shaders. My profile has the same 2600 shaders in fixed share. But on the T4, Cyberpunk 2077 was running at 60 FPS, and I was shocked. Are vGPU profiles significantly weaker in performance?
Also, I saw an old NVIDIA Quadro RTX 6000 (4k shaders). I entered a part of its profile and everything worked well in CyberPunk 2077 there too. It was with vgpu …
What should I do in this case? How can I completely disable FRL so that it never comes back? I read that it’s just a test flag and should not be touched.
You cannot compare the number of CUDA cores between different GPU generations.
To be honest I don’t see a solution to your use case (expectation). You simply won’t be able to run multiple AAA games in parallel on a single GPU with constant 60 fps if you can easily fully leverage the A6000 with a single session.
vGPU is designed for professional applications where you don’t need 100% GPU performance on all sessions at the same time slot (1ms), so the best effort scheduler is very effective. This won’t work for your use case so you may end up with only 2 parallel sessions where you can guarantee the 60fps for AAA games.
Disable FRL should be permanent. Which hypervisor do you use?
Really depends. Equal should be used instead of fixed here. Equal or fixed guarantees the time slice. But equal depends on the number of running VMs so is much better for you.
For professional use, 98% of our customers are running best effort. Indeed, it might help to run with equal and tweak the time slice interval.
Unfortunately. I’m not very familiar with KVM. You may create a support ticket with our NVES to ask for the proper way to disable FRL permanently on KVM although it is not necessary at all when using equal or fixed share as you don’t have FRL active there.
Indeed, as vGPU customer you can only rely on the branch drivers available in the portal. Cloud gaming is a different story. Our cadence for vGPU drivers is much slower compared to cloud gaming but cloud gaming is nothing you can buy easily.
Oh, that’s sad to hear that the drivers cut the performance. Could I find out approximately by how much? I’m not asking for an exact number, but is it 10-20%? What’s the range? I would be grateful for an answer, it would really help me a lot.
Sorry for being unprecise. There shouldn’t be a real difference in performance. I just meant the driver release cadence is different. Gaming drivers are released very often whereas professional driver branches are released less frequent.
For sure it can happen with new games that the optimization is only present in the latest gaming driver that is not available for the professional branch yet. The next vGPU major release will contain the latest driver branch and will be released soon.
You can even change the profile settings in the professional driver so that it should work identical to the gaming driver: