not working: make SLI for VR active suggestion

[Dutch2]

I know only a few use VR and I guess it does not bring extra buyers, but maybe this all is not that labour and cost intensive. 

Is it not possible to make SLI active for BoX as can be read here https://developer.nvidia.com/vrworks/graphics/vrsli

Think members that have an Vcard just at the low VR level would be very thankfull and can go for a cheap sec.hand (980/1070/1080 ranges)  and the highend players that now can finally run the 2080ti or Titan-RTX in SLI mode. 

 

 

Edited February 17, 2019 by Dutch2

[Boundless_I]

Hopefully, developers would find time to bring this stuff into the game. But I still wonder if they are trying to optimise the game better for vr now to make this experience even more immersive then it is?

[Jason_Williams]

I'm told by our Lead Programmer that VR with SLI works. What proof do you have that it doesn't. Is framerate not improved with two cards?

 

Jason

[SharpeXB]

My experience running SLI (not in VR) was that it helped in situations that were GPU limited. In scenarios that were CPU limited it did not. Isn’t much of what limits VR since it’s in 3D, the need to draw two images, which is CPU bound?

[TUS_Samuel]

Look at this table

https://docs.google.com/spreadsheets/d/1gJmnz_nVxI6_dG_UYNCCpZVK2-f8NBy-y1gia77Hu_k/edit#gid=1870565807

 

Il-2 is effectively single-threaded. There is a strong correlation between CPU single thread performance and FPS. Increase in core count or GPU power does not lead to any changes in FPS.

So SLI VR is useless at this point, you can just get 1080 which will be loaded to a half. We need better multithreading.

 

Some clarifications:

- this table has more data for older version. There were some optimizations in 3.00x, but still il-2 is limited by CPU single thread performance

- of course there are multiple threads in il-2, but there is one busy thread and many auxilliary threads which sleep most of the time. So increase in core count actually does give a very small FPS improvement

 

More details in this thread

 

[SharpeXB]

11 hours ago, TUS_Samuel said:

 

Il-2 is effectively single-threaded. There is a strong correlation between CPU single thread performance and FPS. Increase in core count or GPU power does not lead to any changes in FPS.

 

It depends on the situation. In scenarios with many other aircraft or AI, yes the CPU will limit your frame rate and there will be no performance difference from GPUs in SLI. 

In situations like a free flight then SLI will give you a boost in FPS. However we really are concerned with worst case performance so the CPU tends to be the limiting factor. 

When running tests like the one above you need to be aware IL-2 gives you different (better) performance when running a track vs live gameplay. I imagine because the AI or FM isn’t running and you’re just looking at a recording. It’s about a 10fps difference last time I checked but it’s been a while since I looked at it. 

[Dutch2]

Thats a complete other discussion, here it is about is SLI working in VR and nothing more, which is in VR complete different btw. 

 

We do need an SLI user who can test SLI in VR and analyse the FPS improvements he should see a big improvement. 

[Hellequin13]

SLI/VR user here with a Vive and a pair of EVGA gtx 1080 FTWs. Just ran some tests to see how the multi-GPU support works, but ran into a problem straight off: I cannot toggle off the multi-GPU option. I tried through system settings, which resulted in a restart, only to find the option toggled on again. I subsequently tried disabling it in the Setup.CFG file (mgpu_compatible = 0) and making the file read only... the option still shows as enabled. I then tried forcing the game to run on a single GPU (via nVidia Control Panel) and still could not get the option to be toggled off.

 

Regardless of this, I tried a few tests with the cards set to run in (nVidia Recommended) SLI, alternate frame rendering, and single GPU modes. I loaded up a QMB (Kuban summer) with the plane parked on the tarmac, noting the frames as I looked around me. The average frame was ~15ms, which would jump to 20ms when I looked in a new direction, settling back to 15ms once the objects/textures had loaded into memory. Subsequent returns to the same direction would jump to 18ms before settling back to 15. This was true for all but one setting; the nVidia recommended SLI setting seems to struggle when looking towards the water, and was fluctuating between 18 and 20ms.

 

 

[coconut]

To try it in VR, you have to make sure the bottleneck is the GPU first, for instance by setting a high super-sample value. Otherwise one is likely to be limited by the CPU, and any benefits of SLI would go unnoticed.

[Hellequin13]

4 minutes ago, coconut said:

To try it in VR, you have to make sure the bottleneck is the GPU first, for instance by setting a high super-sample value. Otherwise one is likely to be limited by the CPU, and any benefits of SLI would go unnoticed.

 

I am currently running with a 1.8 SS value (in SteamVR) but have run up to 3.2 on Ultra settings. Frames were sub par for game play, but it was not a slide show, as I expected it to be. I am quite certain the limiting factor is the CPU ( mine being a 8370 OC'd to 4.8), but all of this is kind of irrelevant as we are trying to determine if the Muli-GPU option has an affect on performance, and at present there is no way to turn the option off to confirm. I could run the tests again with higher graphical and SS settings but we still wouldn't know if the engine is utilizing the cards efficiently or not.

 

Based on my testing with the different modes, I would say there is not a significant affect as running in single GPU mode ran at the same level as the alternate frame methods, and slightly better than the standard SLI (at least in regards to rendering the water). But how much the engine's implementation of SLI has to do with that is unknown because I cannot remove it from the equation.

 

Some developer input is needed here. 

[Dutch2]

 

Set everthing about graphics to high/ultra. Uncheck all ground activities and avoid to interfere other planes in your flight, so your CPU is free of AI calculations. 

I hope now you will see more FPS drops on the single card. 

 

Edited February 16, 2019 by Dutch2

[Hellequin13]

Alright, so I set the game as follows:

 

Ultra setting with Shadows = Ultra, Mirrors = Complex, Distant Landscape = X4, Horizon Draw = 150km, Landscape Filter = Off, Grass = Ultra, Clouds = High, SSAO = Off, HDR = On, Sharpen = Off, 4K Textures = On

 

NOTE: I still am unable to get the Multi GPU option to toggle off, so all testing is with the option enabled.

 

I then set up a QMB with a P-39 on the Kuban Summer map, parked at the airfield midway down the coast, south east on the map (sorry, I use a track ball ...for precision clicking when 3D modeling... which does not have a scroll wheel, so I cannot zoom the map and therefore do not know the name of the field or village nearby). I set the time to 12:00pm, no wind, turbulence, or clouds. And finally, no ground targets or AI of any kind.

 

I then tested under no supersampling (100%), 200% SS, and 400% SS, switching between SLI, single card, and forced alternate rending methods 1 and 2. Each test consisted of starting the mission, un-pause, and do two sweeps looking at my 9 (south towards water), 12 (west towards tree covered hill), and 3 O'clock (north towards the village). The first sweep was to let everything load, while I would hold my gaze on the second sweep to monitor frame rates.

 

With no SS, SLI and single GPU methods were identical, giving 14-15ms @ my 9, 15-16ms @ 12, and 16-17ms @ 3. GPU usage was ~50% on GPU0, and 5% on GPU1. The Alternate 1 and 2 render methods yielded slight longer frames with 15-16ms, 16-17ms, and 17-18ms respectively. Oddly GPU0 usage went up to ~65% with Alt 2, but was at ~50% with Alt 1. GPU1 was stable at 5% throughout all four tests (yes, even with single card rendering).

 

With SS at 200% SLI and single GPU gave 20-21ms, 21-22ms, and 22-23ms. ALT 1 and 2 gave 21-22ms, 22-23ms, and 23-24ms. Here I saw GPU0 usage up at ~95%, except under single GPU rendering where it was only ~85%. GPU1 stayed consistently at 5% again.

 

At 400% SS, SLI and single GPU gave 26-27ms, 28-29ms, and 29-30ms. Alt 1 and 2 yielded the same frame rates, and GPU usage dropped to ~85% on GPU0 and 4% on GPU1, with SLI rendering being the anomaly with a ~90% on GPU0. I suspect there was some throttling due to temps at this point.

 

My takeaway: Due to not being able to toggle off the Multi GPU option, there is a negligible difference between SLI and single GPU usage. This is very odd, as I would suspect forcing nvidia to use one card should override the game's setting, but this does not appear to be the case. Trying to force the the system to render alternate frames between the two cards seems to be fighting with the game's Multi GPU support, thus the slight decrease in frame rate. And the fact that GPU1 does not exceed 5% usage tells me that the Multi GPU support is not really all that efficient.

 

I am not sure how hard it would be to shoehorn the VRWorks API into the engine here, but I would love to see my second GPU getting put to good use. I know it won't help with the issues caused by the CPU bottleneck, but being able to run at higher SS and graphics settings would go a long way to improve the visual clarity. While I can spot out to ~7km, IDing is quite difficult beyond ~1k. I've lost SA/advantage on more occasions than I care to count because I am spending too much time trying to figure out what I am looking at.

[Dutch2]

On 1/31/2019 at 3:59 AM, Jason_Williams said:

I'm told by our Lead Programmer that VR with SLI works. What proof do you have that it doesn't. Is framerate not improved with two cards?

 

Jason

 

Jason, Think it is not working?  

Please, Let your lead programmer read this: https://developer.nvidia.com/vrworks/graphics/vrsli

 

edit, to be sure, we talking about SLI for VR, not the SLI for the flat monitor world. 

Edited February 17, 2019 by Dutch2

[TUS_Samuel]

Quote

With the OpenGL multicast extension, it’s possible to upload the same scene to two different GPUs and render it from two different viewpoints with a single OpenGL rendering stream. This distributes the rendering workload across two GPUs and eliminates the CPU overhead of sending the rendering commands twice, providing a simple way to achieve substantial speedup.

 

This technology not only utilizes the second GPU but may help to eliminate current CPU bottleneck.

[Field-Ops]

Man I man, every month or so I go back to thinking about the possibility of SLI in combination with VR. So few games support it but this game and how demanding it is would benefit greatly from SLI + VR support. 

 

https://developer.nvidia.com/vrworks/graphics/vrsli

 

 

Sadly so few games have implemented it that I fear the setup I invested in will never come full bloom into what I imagined. 

[Cavalier]

Has there been any update on SLI for VR? does/will it work nowadays?