DirectX9 Fixed-Function Pipeline Performance Regression on 50-series GPUs

Gaming and Visualization Technologies General Topics and Other SDKs Drivers - Linux, Windows, MacOS
1

We’ve observed a performance regression with the 50-series GPUs when running a DX9 application, specifically related to drawing Shadow Volumes via a Fixed-Function approach.

The issue manifests in frametime spikes and is reproducible on a 5060Ti, 5070Ti or 5080. If we swap out the GPU for a 4060Ti, A4500 ADA or A6000 the issue is not present and frametimes are stable.

We’ve tried drivers from v576.02 - 576.52, and are able to confirm that the issue only appears to manifest on the 50-series GPUs.

Currently this issue seems to be related to the fill-rate of the stencil buffer when using fixed-function pipeline as we’ve noticed that the issue scales with resolution, but only with the 50-series cards.

We’re unable to determine why this is the case for the 5000 series cards, when they should perform better than previous cards.

DxDiag.txt (95.7 KB)

2

Hi there @tgreen2, welcome to the NVIDIA developer forums.

I think before I can forward this as an actual bug to Engineering, I would need a bit more detail.

Did you try debugging your application for example with NSight?

Some form of game or graphics capture through NSight would also be necessary in case we pass this on.
Or a minimal app to reproduce it in-house.

Thanks!

3

Hi @MarkusHoHo,

We did collect NSight captures on both the 4060Ti and 5060Ti. As the files are large and we’d prefer they weren’t publicly available on the forums, could you please provide a way for us to send them to you directly?
The 4060Ti capture is ~130MB and the 5060Ti one is ~88.2MB.

Thank you!

4

Hi again

We’ve managed to reproduce this with a sample application and compared a 3060 to a 5060Ti with the following results:
- 5060TI: 2350FPS, 3060: 2800FPS (Window 606x479 with text disabled, setting LOD to 0) - Executable named Shadows.exe
- 5060Ti: 900FPS, 3060: 1300FPS (Fullscreen no text, 3840x2160, setting LOD to 0) - Executable named Shadows_fullscreen.exe

Obviously on our more complicated application, the performance impact is significant enough to not meet our requirement of 60FPS on the 5060Ti, but the 3060 is performant enough.

Attached a zipped version of the app.

Cheers

SampleAppSourceCode.zip (7.2 MB)