EXPERIMENTAL: Resizable BAR in DCS for NVIDIA RTX 30 Series GPUs

[Woona]

43 minutes ago, imacken said:

That is going well above anything anyone else is seeing!

Do a Google search. Values between 1 and 4% are typical, ignoring Nvidia’s site of course!

 

There's a perfectly fine chance that my benchmark situations are just coincidentally stressing my GPU memory sufficiently that the full access to the vmem is beneficial. If that is the case in every situation of mine though, I'm going to the gas station and picking up a lottery ticket, because I just ran another benchmark with ReBAR enabled and then disabled at a BIOS-level - 11,92% uplift in minimums, 11,57% uplift in 99th percentile. Validated replicable benchmark track, fast taxi at McCarran. 

 

I've attached the benchmark TRK. Requires F-16C. 
I begin and end my benchmark with the clock on the DED to make sure I start and stop the exact same place, within milliseconds of each other. I benchmark with Nvidia FrameView. If you wanna give it a shot yourself and see if you have the same uplifts, that'd make for some really interesting data. Otherwise you can create your own fast taxi across a runway at McCarran in whatever plane you please I suppose, as long as it's replicable run-to-run with the same settings, including ReBAR status.

NVCP AniTro Filt test.trk

Edited April 11, 2021 by Woona

[speed-of-heat]

hi there, one thing i dont understand is "where the improvement comes from"... e.g. is the improvement in CPU or GPU frame times or both? 

[Woona]

@speed-of-heat Do you mean that you're not sure how ReBAR functions? I actually haven't measured frametimes, so I don't wanna throw any assumptions out there in that regard.

[speed-of-heat]

nah i know how its meant to work, i meant is the frame time reduced on cpu or gpu or both and if so by how much  

Edited April 12, 2021 by speed-of-heat

[ShaunX]

Ok, I installed NVIDIA FrameView and did some tests.

I created a benchmark track based on FA-18C->Instant Action->Syria->Free Flight. and recorded the first 120 seconds. My only control input was to set baro alt hold.

The track flies with a wingman, from the Turkish coast, North East along the Orontes River valley for 120 seconds at 1500 ft barometric and just crosses the city of Antakya

 

I set Frameview to record 119 seconds.

 

I did my tests in 2D, my spec is in my sig and my settings pretty much match S-O-H's for VR and I have CAS installed

 

I did one run without ReBAR, then 3 with ReBAR enabled via profileinspector , then another 2 without it.

In each case I rebooted first, launched NVIDIA FrameView, Launched DCS, replay the track, then pressed the start record hot key immediately after clicking "fly"

 

Doesn't look like any kind of improvement to me, though I don't really know how to interpret these figures. Summary file attached.FrameView_Summary.xlsx

 

[speed-of-heat]

no actual frametimes in your data looks like you are just recording the percentage numbers from 2d... , but, based on what you are seeing there is no real difference in framerate either

[ShaunX]

No option in Frameview to record frametimes. At least not in the summary file. It also generates a detail file for each run which for my 119 second samples each contains around 21,000 records !

 

There is a column in those files called "MsBetweenPresents", which may be equivilent ?

 

 

..but again, doesn't look like much different between files.

 

The utility also comes with an Excel template that you can plug multiple detail files into to get graphical comparisons. Haven't had a chance to play with that yet but looks useful. Looks pretty good for a free utility.

 

 

[Loki 1-1]

20 minutes ago, ShaunX said:

Ok, I installed NVIDIA FrameView and did some tests.

I created a benchmark track based on FA-18C->Instant Action->Syria->Free Flight. and recorded the first 120 seconds. My only control input was to set baro alt hold.

The track flies with a wingman, from the Turkish coast, North East along the Orontes River valley for 120 seconds at 1500 ft barometric and just crosses the city of Antakya

 

I set Frameview to record 119 seconds.

 

I did my tests in 2D, my spec is in my sig and my settings pretty much match S-O-H's for VR and I have CAS installed

 

I did one run without ReBAR, then 3 with ReBAR enabled via profileinspector , then another 2 without it.

In each case I rebooted first, launched NVIDIA FrameView, Launched DCS, replay the track, then pressed the start record hot key immediately after clicking "fly"

 

Doesn't look like any kind of improvement to me, though I don't really know how to interpret these figures. Summary file attached.FrameView_Summary.xlsx

 

 

Perhaps the fact that the 3090 has 24 gb ram memory vs 10 for the 3080 has an affect on the results. I have the 3080 aswell and will do some VR benchmarks this week. Will post back when done.

[speed-of-heat]

26 minutes ago, ShaunX said:

No option in Frameview to record frametimes. At least not in the summary file. It also generates a detail file for each run which for my 119 second samples each contains around 21,000 records !

 

There is a column in those files called "MsBetweenPresents", which may be equivilent ?

 

 

..but again, doesn't look like much different between files.

 

The utility also comes with an Excel template that you can plug multiple detail files into to get graphical comparisons. Haven't had a chance to play with that yet but looks useful. Looks pretty good for a free utility.

 

 

thats a composite number of both the GPU and CPU, FPSvr can differentiate

[Woona]

 

4 hours ago, ShaunX said:

No option in Frameview to record frametimes.

 

 

If you want to measure VR frametimes, you're looking for OCAT (see https://gpuopen.com/ocat/). It'll give you frametimes in averages and percentiles. It's what Nvidia's FCAT built on before its untimely death. OCAT is a fantastic tool, I just have a preference towards FrameView. Just two unrelated benches here as an example, these have nothing to do with ReBAR.

 

 

I've got some time to mess around again, so I'll go ahead and make a couple frametime measurements for you here shortly, @speed-of-heat .

Edited April 13, 2021 by Woona

[speed-of-heat]

Thanks! @Woona

[Woona]

OCAT frametime validation.
Edit: As Speed has pointed out, these values are in fact in-line with a 7-8% delta as measured on previous runs of this benchmark with Nvidia FV.

Edited April 15, 2021 by Woona

[speed-of-heat]

thats strange!  I mean math says frametimes fall frame rate increase  but 1ms is a about an 7-8% decrease which would tally ish with the numbers you reported

[Woona]

5 minutes ago, speed-of-heat said:

thats strange!  I mean math says frametimes fall frame rate increase  but 1ms is a about an 7-8% decrease which would tally ish with the numbers you reported

Oh shoot, you're totally right. I guess it's about right then.

[speed-of-heat]

whats interesting is that it is across GPU and CPU frametime that have been made more efficient 

[zildac]

I'd also be interested in what difference MB BIOS versions may make to potential gains. As an example, on my Maximus XII the current non beta BIOS which enabled ReBar says this:

 

 

The current BETA BIOS (which I haven't taken the plunge with as yet) says this:

 

[speed-of-heat]

the forum says end of april beginning of may for rebar on asus 370/90

[ShaunX]

17 hours ago, Loki 1-1 said:

 

Perhaps the fact that the 3090 has 24 gb ram memory vs 10 for the 3080 has an affect on the results. I have the 3080 aswell and will do some VR benchmarks this week. Will post back when done.

 

Interesting point perhaps ?

 

Perhaps the 3090's that have been tested in this thread aren't seeing a benefit because they don't need the additional bandwidth as they're already exploiting VRAM well in excess of the 10Gb of a 3080 in these benchmark scenarios.

 

Perhaps we need to test with benchmarks which involve more dynamic environments to see a benefit ?

 

[ShaunX]

4 hours ago, ShaunX said:

 

Interesting point perhaps ?

 

Perhaps the 3090's that have been tested in this thread aren't seeing a benefit because they don't need the additional bandwidth as they're already exploiting VRAM well in excess of the 10Gb of a 3080 in these benchmark scenarios.

 

Perhaps we need to test with benchmarks which involve more dynamic environments to see a benefit ?

 

 

..There may be something in this theory.

 

I created a benchmark track with a 2 minute, high speed, NOE, high G turning circle in the Viper over the city of Damascus - Syria NOE Tight Turn Benchmark.trk

 

I did 3 tests without ReBAR and then 3 with, rebooting between each test

 

..now I get a consistantly higher average FPS

 

Interestingly my CPU and GPU temps are also consistantly lower.

 

[Woona]

@ShaunX More data is always good, thanks! That's a 14,05% uplift (121 -> 138) in 99th percentiles.

I'm also experiencing lower temperatures on my GPU and CPU, interestingly. I thought it was just a freak coincidence every time but if it's happening to others as well, I guess there's definitely some increased efficiency going on here.

 

I'm starting to ponder if this should go in the 2.7 Optimization Guide. I just also don't want even a single case from a poor guy who bricked a single-VBIOS 3000-series card or whatever. That'd break my heart. But there's obviously performance to be had, so far as we can measure, ReBAR or not (until someone provides a way to check on a technical level if ReBAR is truly working, anyway). Edit: Speaking of which: I heard from LTT recently that Nvidia told him ReBAR is simply an enabled profile for the individual games. Unless they're specifically hiding technical details, this method of using NPI and enabling ReBAR flags would make perfect sense. But alas, I don't have a way of actually verifying other than our measurements.

Edited April 14, 2021 by Woona

[ShaunX]

It seems our 3 flag method is making its way into other publications..

 

https://wccftech.com/heres-how-you-can-enable-resizable-bar-support-in-any-game-via-nvidia-inspector/

 

https://babeltechreviews.com/z390-resizable-bar-performance/

 

 

 

 

[Woona]

15 minutes ago, ShaunX said:

It seems our 3 flag method is making its way into other publications..

 

 

Here's hoping Nvidia doesn't make it lock it down to something at a lower level, but listens to their customers and gives us the option to manually enable it for individual games. Maybe nothing changes at all, that'd be OK too.

[speed-of-heat]

just like they did with VRSS you mean ...

 

Edited April 14, 2021 by speed-of-heat

[Buzzer1977]

Have you verified that Resizable Bar is not only enabled but used ?
To check if your system has BAR enabled (Supported CPU, BIOS, BIOS settings, VBIOS, driver, ...)
open the NVIdia control panel (right click on the desktop) and check the system informations in the help menue.
 

To see if the system is actually using BAR open the GPU view in the task-manager and watch the shared memory segment while running the application.
0.1 GB is standard on a idle system and no sign of actally using BAR in the application.

 

[ShaunX]

23 minutes ago, Buzzer1977 said:

Have you verified that Resizable Bar is not only enabled but used ?
To check if your system has BAR enabled (Supported CPU, BIOS, BIOS settings, VBIOS, driver, ...)
open the NVIdia control panel (right click on the desktop) and check the system informations in the help menue.
 

To see if the system is actually using BAR open the GPU view in the task-manager and watch the shared memory segment while running the application.
0.1 GB is standard on a idle system and no sign of actally using BAR in the application.

 

 

 

We all went through the steps to enable and verify it's enabled in NVIDIA system information

 

However, if you're correct then my benchmark is NOT using it as shared gpu memory remains at 0.1 GB throughout.

 

Is there an article which describes that behavoir ?