DCS *in pancake mode* prefers DDR3

[TAW_Impalor]

Hello. As some of you may know, I have run a few tests to see performance difference between 2.7.6 vs 2.7.7 in this thread. My loss was somewhat smaller than what many others observed. I think I finally know why!

Today I ran the same tests to compare performance of my new 12th gen rig vs old 4th gen, using the latest OB 2.7.8. I went crazy and bought an expensive kit of some of the fastest DDR4 for the build: 4000cl14. Unfortunately, Aorus Elite only allowed me to run it at 3466cl14 due to BIOS bugs. I kept 1080ti until I can buy the 4090. My old PC was running close to 100% CPU, GPU, RAM (16Gb) and VRAM (11Gb) utilization in multiplayer while in VR, and gave me about 45 FPS, but there was not enough CPU cycles left for motion smoothing. Head tracking was also jittery, so I decided it's time to upgrade.   

Anyway, I started with 1080p mode, and to my surprise and disgust the new rig performed MUCH WORSE than the old (-58% difference)! I noticed that it utilized little CPU and GPU. Everything was optimized, except CPU overclock, as benchmarks showed. No HT, no E-cores.

On a hunch I assumed the memory was too fast (haha) and tried decreasing the clock. Guess what? CPU and GPU utilization, and FPS, increased! I went all the way down to 1066 (equivalent to my old DDR3-2133), to get the best results. But they were still short of my old rig!!! Below 1066 clock the new mobo did not boot (why??). cl14 timings were also the best it could post.

Next, I tested in VR, and the relationship was the opposite - faster RAM gave better results, but overall gain over the old rig did not exceed 6.9%. I understand and accept this, because GPU is still a bottleneck for now (and my DCS was optimized to run it at full capacity before), however it made almost no difference between 3466 and 2133 RAM clock. 

The test track involves descending through overcast (8v8 F18 mission), so I think it is the clouds making gen 4 rigs perform better or similar to gen 12. Of course, I used very low settings that let me play in VR (no shadows, no AA, low preload radius). The detailed test results and settings are here, with the summary below (I ran each test twice for accuracy):

Hopefully, this can help the devs in their analysis why DCS performance is so poor. As always, YMMV.

Edited November 25, 2021 by impalor

[Flappie]

Please ensure your Windows power management is set to the max.

[TAW_Impalor]

2 hours ago, Flappie said:

Please ensure your Windows power management is set to the max.

 

This? I did not reinstall Windows, all settings remained the same from 4670k build. Just flew VR on GS server, FPS was no higher than before (40-45), but experience is smoother, no stutters when I am low and turn my head to check 6. I saw 22/32 Gb RAM used, 10.5/11 VRAM, GPU utilization 100%, CPU 40% (8 P-cores). Some say my settings are "Minecraft", but for PvP that's ok.

[Flappie]

Yes, that's what I was talking about. I have no clue why you're getting such a difference between VR and pancake in 2.7.8. Have you done the same test in 2.7.6?

Edited November 24, 2021 by Flappie

[TAW_Impalor]

15 hours ago, Flappie said:

Yes, that's what I was talking about. I have no clue why you're getting such a difference between VR and pancake in 2.7.8. Have you done the same test in 2.7.6?

 

Updated my analysis to include 2.7.8 vs 2.7.6 on DDR4-3466 vs 1066. The results are shocking: 

My conclusions:

1. Pancake guys with good CPU/GPU should stay on DDR3 or downclock DDR4.
2. OB 2.7.8 benefits from downclocking more than 2.7.6, but they both do.
3. VR guys benefit from DDR4, but buying ultra fast kits does not matter.
4. The better the rig the higher was the FPS loss going from 2.7.6 to 2.7.8. 
 

Edited November 25, 2021 by impalor

[Limaro]

22 hours ago, Flappie said:

Please ensure your Windows power management is set to the max.

I saw btw. that my settings were only to "Ryzen Balanced" but also on MaxPower, leads to identical results.

@impalor I tried a lower multiplier for the RAM, down from 3200 to 1600. but this had no effect on my side.

[TAW_Impalor]

11 minutes ago, Limaro said:

I saw btw. that my settings were only to "Ryzen Balanced" but also on MaxPower, leads to identical results.

@impalor I tried a lower multiplier for the RAM, down from 3200 to 1600. but this had no effect on my side.

Was your GPU utilization already high?

[Boosterdog]

Have the same tests been run with reasonable 2D Graphics settings or just the bare bones one previously used? 

[Hiob]

Can't reproduce.

Ryzen 5900X/RTX3080 in 4K

When disabling XMP (3600 -> 2133), I loose between 3-5% FPS - as expected I would say.

[TAW_Impalor]

2 hours ago, Boosterdog said:

Have the same tests been run with reasonable 2D Graphics settings or just the bare bones one previously used? 

Same. 60 fps do not suggest I can improve visuals. Why barebone settings run so slow with fast memory, but double the FPS with slow? That is the question. I only fly VR, so the test was rather theoretical.

[TAW_Impalor]

2 hours ago, Hiob said:

Can't reproduce.

Ryzen 5900X/RTX3080 in 4K

When disabling XMP (3600 -> 2133), I loose between 3-5% FPS - as expected I would say.

Hello. To reproduce, did you run with my settings (Preset 1 in the Google drive), or yours?

[Boosterdog]

10 hours ago, impalor said:

Same. 60 fps do not suggest I can improve visuals. Why barebone settings run so slow with fast memory, but double the FPS with slow? That is the question. I only fly VR, so the test was rather theoretical.

Except they don’t in general. Only it seems for you. From the previous runs I did with you setting albeit a 1920 1080 resolution to match my monitors native resolution  I saw an average of 116 in 2.7.8 v 124 in 2.7.6.  
 

taking into account resolution and your higher spec gpu to mine I’d say that’s not too different from your “discovery”  results in this test. And my results were obtained on DDR4 @ 3600. 
 

that’s why I asked if the test had been tried using more realistic 2d graphics settings to actually put some pressure on the system at VR does.  
 

I don’t think you have sufficient corroborating evidence to make the claims you are making let alone self declare it as a solution at this time  

 

[Hiob]

9 hours ago, impalor said:

Hello. To reproduce, did you run with my settings (Preset 1 in the Google drive), or yours?

Hi, no I just changed the XMP-Profile. I used the same (rather high) settings, I use most of the time. So I'm definetely GPU bound in the first place.

I assume, your findings are an edge case, that are only relevant with very CPU-bound scenarios... and or that you have found a very specific hardware bottleneck. Interesting nevertheless.

[TAW_Impalor]

5 hours ago, Boosterdog said:

Except they don’t in general. Only it seems for you. From the previous runs I did with you setting albeit a 1920 1080 resolution to match my monitors native resolution  I saw an average of 116 in 2.7.8 v 124 in 2.7.6.  
 

taking into account resolution and your higher spec gpu to mine I’d say that’s not too different from your “discovery”  results in this test. And my results were obtained on DDR4 @ 3600. 
 

that’s why I asked if the test had been tried using more realistic 2d graphics settings to actually put some pressure on the system at VR does.  
 

I don’t think you have sufficient corroborating evidence to make the claims you are making let alone self declare it as a solution at this time  

 

I don't think you understood my results. In VR, there GPU is fully utilized, lower clocked memory is detrimental. I think increasing 'eye candy' in 2d will have similar effect. It is with 'barebones' settings in 2d I see a surprising effect that CPU and GPU are idling! And to combat that, on my PC at least, downclocking memory all the way to 1066 (half the SPD) push the performance higher.

[Boosterdog]

2 minutes ago, impalor said:

I don't think you understood my results. In VR, there GPU is fully utilized, lower clocked memory is detrimental. I think increasing 'eye candy' in 2d will have similar effect. It is with 'barebones' settings in 2d I see a surprising effect that CPU and GPU are idling! And to combat that, on my PC at least, downclocking memory all the way to 1066 (half the SPD) push the performance higher.

No I get it. I just dont agree with your ascertion that DD3 is better for pancake as most pancakers dont have such stripped back settings and are hammeing their GPUs every bit as hard as Gavin in the Goggles. You appear to be missing or ignoring that fact in making the headline statement. 

[Hiob]

In science there is very high valued principle that goes "the bigger the claim, the better/more evidence is needed to proof it" or „extra ordinary claims need extra ordinary proof“.

I think, something similar is, what @Boosterdog is trying to tell you. The claim is pretty big, because it goes completely against the common experiences regarding memory speed. So there is definetely need for further evidence, before you can call it a day.

 

Edited November 26, 2021 by Hiob

[TAW_Impalor]

It is a fact that upgrading from 4670k/16Gb DDR3/1080ti to 12900k/32Gb DDR4/1080ti my FPS dropped 50% on 1080p screen. It increased 6% in VR on the same settings. I furter observed inverse relationship between memory bandwidth and FPS on my PC in 1080p. I have no motive to lie about that. I informed the commutiny and devs about this strange fact in hope that it would help debugging the decrease in performance from version to version.

[Boosterdog]

@impalor. I wasnt accusing you of being dishonest. If that came across I sincerely apologise. I mean only that the basic ascertion that 2D works better with slower DDR4 RAM (I cannot test DDR3 is not true in most instances. It really only applies, it appears, to a very specific set of circumstances and settings that that average 2D user would not reduce their settings to. 

This is not to say that I dismiss your findings. Quite the contrary. They are certainly intersting results and something that I think should be looked at to work out why it occurs. 

Again, if I have offended you I apologise. 

[okopanja]

I would suggest to check your temperature sensors for GPU and CPU. I had similar issue with thermal throttling, where lowering my CPU max speed helped me achieve higher frame rates. In my case it was a laptop, with GPU and CPU sharing the same heat-sink. When the CPU went up, it heated the GPU, and caused it to go into thermal throttling.

Use nvidia tools GeForce Experience or Frameview to inspect what is going on. HWMonitor can be also used to see the peak values.

https://www.nvidia.com/en-gb/geforce/technologies/frameview/

Both have OSDs which you can use to track what is going on, while playing...

If indeed temperature is the one causing the issue, possible solution is to ensure your fans runs at maximal speed, when running DCS.

 

Edited November 29, 2021 by okopanja

[QuiGon]

Interestring thread. I'm curently running a Gen 4 setup (i7-4790K, 16 GB DDR3, RTX 2080) but I already own a Gen12 MoBo and CPU (i7-12700k) which I can't run yet as I'm still missing the memory for it (32GB DDR5).

I play DCS in pancake mode and have maxed out the visuals. It runs smooth unless I'm playing on a busy multiplayer server with lots of players and AI units (BlueFlag server) which my RAM (or my Windows page file?) struggles to keep up with. I'm curious to see how my updated rig will perform.

Edited November 29, 2021 by QuiGon

[okopanja]

23 hours ago, QuiGon said:

pancake mode

maybe I am just too new and this is kind off-topic: what is pancake mode?

[Sile]

Zitat

maybe I am just too new and this is kind off-topic: what is pancake mode?

non-VR, i guess

Edited November 30, 2021 by Sile

[Hiob]

You're right. Pancake mode refers to good old 2D.

[Boosterdog]

1 hour ago, Sile said:

non-VR, i guess

 

 

1 hour ago, Hiob said:

You're right. Pancake mode refers to good old 2D.

Yup theres "pancake mode" and "princess mode"

KIDDIN

[Jayhawk1971]

Princess? Pah! VR is the Mode of Kings!