Some semi-controlled testing of non-GPU limited performance (U13)

[Morbad]

Basic procedure here was to eliminate GPU utilization, but ensure any potentially non-GPU dependent settings were accounted for, by running at maximal settings, but a low enough internal resolution to keep GPU utilization below 90%. In this case, that means 1440p Ultra++(ish) with x0.5 supersampling (720p effective internal res).

Edit: I did find an anomaly with the default ultra directional shadow preset, and only this preset of this setting, where it was inconsistently harming frame rate without always increasing GPU load. Avoid the default ultra shadows and any graphics settings that keep utilization well short of maxxed out should be completely down to the CPU, memory, and driver/API overhead.

System used is my rather carefully tuned 5800X3D box that has a lean Windows install and fast system memory. GPU is an AMD RX 6800 XT with the 22.8.2 drivers. A few more details in the video descriptions.

This first scenario is an AI-light area to focus on the game's renderer itself, rather than it's difficulties managing complex NPC interactions or broken settlement assets. I'll test some more complex scenarios and a broader spectrum of systems later.

D3D:

Source: https://www.youtube.com/watch?v=bFc4O-Fe6BU


Vulkan:

Source: https://www.youtube.com/watch?v=s9kpUg178XE


Edit: See post #15.

If anyone wants to do a comparable test, head to any lightly-trafficked, mediumish wealth, industrial, coriolis starport, stand on the center of the bridge and look at the cluster of terminals while facing the mailslot. The less crap running in the background, the better. Make sure you have no frame rate caps of any kind and aren't GPU limited. If you're using DXVK or a non-Windows OS, specify that. System specs would also be nice, especially CPU, memory, and GPU brand.

More points of comparison are always appreciated.

 

[InDigital]

22.8.2 no orange sidewinder or planetary crash?

 

[Unknown Zombie]

22.8.2 no orange sidewinder or planetary crash?

Only with DXVK (if you're on AMD 6000 series).

 

[DrinkyBird]

Wouldn't dropping graphics settings as low as possible be more ideal for reducing GPU usage, instead of cranking them high? Dropping resolution reduces the amount of pixels to fill but there's still a lot of work behind each pixel, which might even be in off-screen buffers that aren't the same resolution as the display. (Although, I recall some lower settings hurting performance more than higher settings, so who knows.)
And what does "Ultra++(ish)" mean, exactly? Providing your exact settings would be helpful if you want others to contribute more meaningful results.

 

[Morbad]

22.8.2 no orange sidewinder or planetary crash?

22.8.2 behaves the same way as 22.5.2 through 22.8.1.

The orange sidewinders and most other outright crashes can be eliminated by disabling double precision floats in AppConfig.xml. There are still some rendering issues, but those can be mitigated to a large degree by using DXVK.

Wouldn't dropping graphics settings as low as possible be more ideal for reducing GPU usage, instead of cranking them high? Dropping resolution reduces the amount of pixels to fill but there's still a lot of work behind each pixel, which might even be in off-screen buffers that aren't the same resolution as the display. (Although, I recall some lower settings hurting performance more than higher settings, so who knows.)

It might also cut settings that have meaningful CPU or memory load, especially if they are anything related to geometry.

In practice, as long as GPU utilization is never near maxed out, and the card isn't running low on VRAM, most of the graphics settings themselves have almost no impact on frame rate.

And what does "Ultra++(ish)" mean, exactly? Providing your exact settings would be helpful if you want others to contribute more meaningful results.

It's my custom settings, but reverting to the default ultra preset changed almost nothing, so I left them as they were.

 

[Morbad]

@DrinkyBird I double checked the effects of my settings and did find one anomaly. The default ultra qualty directional shadow preset setting behaves very strangely and I've added a note to my OP to avoid it for this test. I saw as high as the ~200 fps I was getting in the videos above and as low as ~160 fps between restarts and setting swaps.

No other setting had any impact on frame rate if GPU utilization was not near maxxed out (e.g. I could induce the effect by manipulating only directional shadow quality, but not by manipulating any or every other setting). My custom ultra+ settings at 1440p x0.5 supersampling was the same as the medium preset at 1080p x0.5 SS, for example.

Probably some unlisted variable that is tied to that shadow preset as my custom shadows look better and have a higher slice size, but also perform much more consistently.

 

[danteuk]

Nice to see a reasonable improvement in FPS using DXVK - it's shame some developers are still ignore Vulkan and we have to use an open source unsupported translation layer to get the benefits of the better API. I'm sure eventually MS will steal the improvements that Vulkan brings and add them to DX.

 

[Morbad]

Nice to see a reasonable improvement in FPS using DXVK

The performance improvement is limited to CPU dependent areas...the native D3D11 is still faster for GPU limited ones. Not sure which I'd prefer if I could get them equal in IQ/stability.

I still have to test the game again after disabling the shader Balvad isolated to see how much performance changes, especially in areas that the visual anomalies were preventing me from comparing previously.

I'm sure eventually MS will steal the improvements that Vulkan brings and add them to DX.

D3D12 and Vulkan are very comparable in a lot of technical aspects and have some common ancestry in Mantle. They all focus on reducing driver overhead, enabling better cross platform support, mutlithreading, and lower-level access to hardware.

 

[danteuk]

I'm curious about how you got DXVK and it's hud working with ED? I've copied in the current .dll into the directory EliteDangerous64.exe and I tried running the edlauncher from a batch file which sets the
set DXVK_HUD=1
but it doesn't show the hud.

 

[Morbad]

I'm curious about how you got DXVK and it's hud working with ED? I've copied in the current .dll into the directory EliteDangerous64.exe and I tried running the edlauncher from a batch file which sets the
set DXVK_HUD=1
but it doesn't show the hud.

I haven't even tried to enable the DXVK HUD.

 

[Jack Winter]

A quick google seems to indicate adding the dxvk.hud option to dxvk.conf is how to do it on windows.

 

[Morbad]

Testing the shader workaround for D3D11 now and performance appears unchanged, while all visual issues related to the new drivers seem resolved.

I can confirm, again, that heavily GPU limited areas are ~10% faster on the current AMD drivers via native D3D11 vs. DXVK. Will evaluate some CZ performance (which usually has a good mix) when I find a nice one.

 

[danteuk]

A quick google seems to indicate adding the dxvk.hud option to dxvk.conf is how to do it on windows.

I googled too and just used:
set DXVK_HUD=devinfo,fps
in my batch file, that gave me the basic DXVK FPS overlay - which was really the only way to know for sure it's using DXVK.
My quick test just walking around a very quiet settlement ( at night ) gave no real difference in FPS. I definitely need to test in a station too and play with settings and see what happens if lower the resolution etc, ideally since I don't have any specific issues I'm only testing out of curiosity.
GPU is RTX 3060TI current drivers ( at time of writing, ie updated today ).

 

[Morbad]

I googled too and just used:
set DXVK_HUD=devinfo,fps
in my batch file, that gave me the basic DXVK FPS overlay - which was really the only way to know for sure it's using DXVK.
My quick test just walking around a very quiet settlement ( at night ) gave no real difference in FPS. I definitely need to test in a station too and play with settings and see what happens if lower the resolution etc, ideally since I don't have any specific issues I'm only testing out of curiosity.
GPU is RTX 3060 current drivers ( at time of writing, ie updated today ).

NVIDIA has long had very well optimized and multi-threaded D3D11 drivers and it's unlikely that DXVK will improve things.

 

[Morbad]

I'm pretty sure I've isolated the performance anomaly I was detecting with the default shadow quality settings to a spot shadow setting I left in my override. My comparison between D3D and Vulkan is representative of the difference between native and DXVK, but it's not comparable to others because of this oversight.

I'll make sure I restore all my files to their unmodified defaults and rerun the tests.