DLDSR reduces input lag (Confirmed by Frameview + video)

[Chief Blur Buster]

GPU power management adds frametime jitter when you underload a GPU. This is a problem when using GPUs with emulators sometimes.

One fix for ultra-low-GPU-utilization situations is to use RTSS Scanline Sync with the Force Flush. This bypasses GPU power management, reducing latency of Present()-to-photons.

[andrelip]

Reviving this topic to add some metrics.

I tried this with Frame view 1.4 and Fortnite.

1920x1080 vs 2880x1620 (DLDSR res)

I tested in Fortnite, setting the render scale to 16% and testing both resolutions.

1920x1080 gave me 13-14ms of System Latency while 2880 gave me 8-9ms.

Video:

Capped at 180fps: https://youtube.com/shorts/UWRmp3WNyrs?feature=share
Uncapped: https://youtube.com/shorts/xKtPmGKz9NQ?feature=share

Considerations:

3080 - GPU frequency was locked and frequent in both cases.
The tool reported a smaller latency in 2880 even with higher GPU usage.

[Melkxz]

That is strange, I used DLDSR on my 1440p monitor for solo game. With 3080 too.

And sometime forgot to put it off before playing multiplayer games. Every time I could tell that I forgot to turn it off because of the added blur to the image and the added input delay.

I use ultimate performance power plan, locked CPU/GPU frequency.

And it was mostly noticed in League of Legend that use something like 5/10% GPU/CPU to reach 200fps locked.

[andrelip]

This new Frameview was made to support DLSS 3 with Frame Generation and is reading from latency markers used by Reflex supported games.

While I don’t discard this unusual flow (DLDSR) to create issues in the measurement, I think it’s unlikely.

When I first wrote this, I was suspecting the behaviour without any type of data.

If you guys can try to reproduce and we confirm that this behaviour is not only in my setup, then we can ask for someone to confirm with button to pixel equipment.

[andrelip]

Windows Registry Editor Version 5.00
[HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\Class\{4d36e968-e325-11ce-bfc1-08002be10318}\0000]
"DisableDynamicPstate"=dword:00000001

and test again

With this command it locked the freq in 1985 instead of the default of 1920 from "prefer high performance" (Asus Strix 3080 OC). Now 1080p seems to have a very small advantage. The device was listed at 0002 instead of 0000.

What is curious it that before it was locked in the afterburner and you can even see that in the video.

[andrelip]

it forces highest p state, just use and re-do, i think the test is a joke

I don't know if you noticed, but I agree with you.

With this command, my fans and temp are a little bit higher on the desktop, and the latency of non-DLDSR actually decreased. Actually, now the latency is scaling with GPU usage, as expected.

What I'm commenting here is that before I use this command, I locked the voltage/frequency curve in Afterburner and it seemed to be locking the p-state.

nvidia-smi reported p0 and all the charts seemed to be constant, including voltage. I work with deep learning and Nvidia-SMI was very reliable for memory reading and throttle, at least in Linux.

Although the key is name "DisableDynamicPstate", I think this switch does more than just p-states. Maybe NVIDIA reused this old key to toggle additional power-savings techniques.

[andrelip]

Ignore my last message, it still shows a significant difference even after this command.

Now I tested in Valorant and I got 3~4ms of system latency using DLDSR against 8~9ms of 1080p.
The mouse does feel better.