Huge performance losses with newer NVIDIA drivers

I just spent about 4h playing around with different driver versions to narrow down why this occurs.

Basically Deus Ex Mankind Divided performs significantly and measurably better on the old 375 branch than it does on the new 38X branch. I haven’t yet tried other games.

I’ve done a lot of testing and also played around with environment variables (__GL_THREADED_OPTIMIZATIONS has no effect, so it’s not that) as well as the Composition Pipeline, V-Sync and basically all driver versions that came after the last 375 one.

Here are the results:

  • 34 FPS in-game and 43.3 FPS avg in benchmark on 387.34 from yesterday
  • 47 FPS in-game and 50.6 FPS avg in benchmark on 375.82 from July

The oldest 38X driver I could get to run was the 384.47 Beta from June and it had virtually the same performance as the new 387.34.

My in-game settings and the rest of my software configuration has remained exactly the same to rule out any other factors.

My system for reference:

  • Intel Core i5 2500 4.1 GHz
  • ASUS GTX 970 STRIX OC 1.28GHz
  • 4x2 GB DDR3
  • Arch Linux 4.13.12-1
  • GNOME 3.26.2-1
  • xorg-server 1.19.5-1
  • 1080p display

So now I’m basically looking for input to troubleshoot, as well as other people’s experience with these versions.
If you have the time you can try out the old 375.82 driver (Linux x64 (AMD64/EM64T) Display Driver | 375.82 | Linux 64-bit | NVIDIA) and compare performance to see if this behavior is reproducible across multiple machines.

Thank you.

Some benchmarks from Phoronix with 375 and 384 driver:
[url]https://www.phoronix.com/scan.php?page=article&item=deus-exmd-3driver&num=2[/url]
[url]https://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-Bigger-Pre-4.15-Comp[/url]
Don’t show regression for Deus Ex at 4k.
Though I can remember that Gnome 3.26 has problems properly unredirecting full screen windows, maybe this has a larger impact on 384+ than 375? So something changed in the driver regarding compositing? Maybe try to run the benchmarks under openbox.

Alright I just did some more testing, here are the results:

  • 35 FPS in-game and 45 FPS avg in benchmark on 387.34 in GNOME
  • 42 FPS in-game and 46 FPS avg in benchmark on 387.34 in Openbox
  • 49 FPS in-game and 44 FPS avg in benchmark on 375.82 in GNOME
  • 49 FPS in-game and 45 FPS avg in benchmark on 375.82 in Openbox

Benchmark performance is definitely within the margin of error, but there still are losses on the new driver, even on Openbox.

I also have tearing both in GNOME and in Openbox, so I don’t think the compositor is interfering.

Any other ideas?

Thanks.

How full is your video memory, as reported by nvidia-smi, while running the game?
Can you confirm that it is full with 387.34, but not full with 375.82?

Interesting, here’s what I’ve found:

CPU usage is also much lower on the 387.34 driver. What’s causing this?

This likely matches a similar performance drop observed on another Feral game, Total War: WARHAMMER.
We’ve been tracking it internally as bug 1963500. There was a change, introduced in our r378 branch, to the logic of allocation of certain textures, but it apparently exposed a bug in our memory manager.
Our next release branch, r390, will carry a workaround, and we’re still working on finding and fixing the root cause.

Really interesting, glad to hear it’s not an issue on my end!

Does it affect all cards or just the Maxwell architecture?

We don’t know the root cause yet, so it’s hard to make a reliable comment. That said, I would expect it to impact all GPUs, at least when there isn’t a very large amount of VRAM.

Can you tell us roughly when the 390 driver will be available?

I believe we may release r390 beta before the end of the year.

Alright great! Until then, is there anything we can do from our side like an environment variable we can use to restore the old behavior?

Hi,

fyi, i notice exactly the same issue with my gtx970 and Feral Game company of hereos 2.
Even medium settings for texture quality let the memory explode :(

Cu,
Christian

There isn’t, in currently released drivers, a way to disable the new logic that is causing the problem.
Please downgrade to a driver version that isn’t affected, until we’ve released a driver with a fix or the new logic disabled by default.
Thanks

Might be interesting if the 970 is especially hit by this due to jumping the 3.5GB barrier.

I’ve noticed increased VRAM usage in X-Plane 11 with the newest drivers as well. The usual settings used to consume around 6-8GB on a 1080Ti, but current 387.xx drivers fill all 11GB.

I also can add Dawn of War 3 to the games with issues. Using Opengl the vram in my card shows 4GB which heavily cripples performance. Using the same graphical preset doing the benchmark in vulkan mode, i only see ~ 2.5gb vram usage and everything is working fine.

Michael from Phoronix has found out that many more games are affected.

Certainly seems like 3.5/4 GB meme problem

What is the last release not affected by this bug?

375.82 from July, that’s what I did my testing with. The entire 380 branch is affected.