DCS 2.9 MT uses the WRONG cores again! AMD Ryzen 9 7950X3D unstable\fluctuating FPS.

[Spectre1986]

As you should know, AMD's new 7950X3D and 7900X3D have two CCD (clusters), with one dedicated to gaming. The gaming CCD (first 8 cores of 7980X3D)) has a massive 96Mb of cache, the other CCD only has 32MB and is reserved for productivity tasks (non-gaming) DCS 2.9 MT is using many cores from both CCDs, potentially causing low FPS and FPS fluctuations far worse than ever seen on previous MT versions.

All games, including OLD non-MT DCS, are automatically isolated to the gaming CCD as they should, which are the first 16 threads in Task Manager for the 7950x3D. The problem is that DCS 2.9 MT (and the new 2.9 Single Threaded) uses many cores from both CCDs. This is wrong. There is a substantial latency penalty when communicating between CCDs that can cause studders. It does NOT matter how minor or insignificant you think these tasks are, simply using ANY cores on the second CCD for ANYTHING causes a latency penalty as the infinity fabric is used. DCS needs to use the first CCD's cores of the top-end AMD x3d CPUs ONLY for all threads.

I believe this is causing these new massive fluctuations in FPS (45-90) that I'm seeing, even in the main menu. You can see this in my screenshot and note that my graphics settings are set to the minimum possible for this main menu screenshot. The main menu always had 90 FPS in previous versions prior to 2.9 and these same fluctuations happen in-sim at any graphics settings. Something needs to be fixed here.

From what I understand, AMD uses the "xbox game bar" to determine if the load is a game or productivity workload. If this is true it seems the Xbox GB detects the old standard DCS as a game, but 2.9 ST and 2.9 MT DCS as a productivity workload.

Unrelated, but many performance metrics seem to be missing, for example I only get 0 for the "Simulation" frame time when pressing RCTRL+Pause twice and expanding the block of data, this frame time does seem to show on the graph though.

Before running this test and capturing this screenshot I cleared the FXO and Metashaers2 folders and also deleted all of my multiplayer tracks. It didn't make a difference.

7950X3d
4090
64GB @ 6Ghz

 

 

Edited October 20, 2023 by Spectre1986

[Spectre1986]

Here is my log file dcs.log (from C:\Users\<youruser>\Saved Games\DCS.openbeta\Logs).
I used the mt.lua file for expanded logging (it makes your log file more useful to the devs when you drop it in C:\Users\<youruser>\Saved Games\DCS.openbeta\Config).
Here is my DxDiag.txt (run > DxDiag, let it run then "Save all information").

Here is a screenshot from in-sim with very low settings, it shows the same repeating fluctuations from 45-90FPS that we get in the main menu. My PC should easily provide 90 FPS with these settings, especially in the menu. You can also see that DCS is loading up most of my non-gaming cores in task manager 16-31.

Please find and squash these bugs, 45 FPS VR looks pretty bad when you're accustomed to 90FPS. The constant fluctuations between the two is really bad.

null

7950X3d
4090
64GB @ 6Ghz

Edited October 20, 2023 by Spectre1986

[Dodge]

2 hours ago, Spectre1986 said:

From what I understand, AMD uses the "xbox game bar" to determine if the load is a game or productivity workload. If this is true it seems the Xbox GB detects the old standard DCS as a game, but 2.9 ST and 2.9 MT DCS as a productivity workload.

While it probably won't make a difference, but to rule out the "xbox game bar" as an issue; check to see whether you have the latest version installed, if you haven't already. I noticed mine updated on the 12/10/2023) when I ran the 'Get Updates' within the "Microsoft Store" program - Win11 user here.

null

 

 

 

[Spectre1986]

My Xbox game bar app did need an update, but updating did not help. DCS 2.9 still uses the wrong cores and you can see the fluctuations that persist from the main menu through to in-sim. The top graph is the menu, the bottom is in-sim with minimum settings. In case you didn't know, you display this graph with RCTrl+Pause. I even tried some heavy underdamping to really reduce the load on my top-end PC, it still fluctuates like the main menu. 

null

Edited October 20, 2023 by Spectre1986

[Luka CMF]

like the code is purposely written in a way that it only fully utilizes intel CPUs.  

[Deanzsyclone]

Same problem here I spent hours on it yesterday. The new update 2.9 I have dropped from 90 FPS down to 60ish, through hours of tweaking I discovered that the biggest culprit is the Shadows setting is overloading my 7950x3d #4 CPU. Turning off Shadows completely significantly improved, and then lastly the next biggest thing that was killing my frame rate but not as bad as Shadows was setting visible range to medium. Now I have my 90 FPS back with no spikes in my frame timing and reliable. But of course I'd like to have my Shadows back and my draw distance out to what it was before. But for now this is a temporary fix I hope somehow they're getting this information to them and will be making a quick fix. 64 gigs of memory, 7950x3d, and a 4090 GPU, using my pimax crystal VR. 

[Spectre1986]

Thanks for the input Deanzsyclone, we have the same config, but my issue seems to start in the main menu and persists thru to in-sim at any settings. Can you load the sim up to the main menu, press RCtrl+Pause to show the frame time graph and get a screenshot ("prt sc" on keyboard) and paste (Ctrl+V) it here? I'd like to see if you have the same fluctuations. I'll also test your shadow settings recommendation and report back.

I just went and did the latest BIOS update for my "Asus Tuf Gaming x670e-plus WiFi" motherboard and then did all available driver updates, including GPU. It made no change for me. It's still loading the wrong cores and giving a terrible experience. I have included the screenshot from the "Huey Hard Ground Attack" quickstart mission on the default map. Note: I'm not loading any extra software (SimShaker, SimAppPro, etc.) for these tests, I'm only running Pimax and DCS.
 

Here is my newest dcs.log file with the expanded logging mentioned previously.

null

[Spectre1986]

I resolved the fluctuating FPS issue by "resetting defaults" in the OpenXR Toolkit via the in-sim Ctrl+F2 menu. I did this when I initially enabled quad views per the instructions, but I think I changed some settings when troubleshooting DLSS visual quality issues and it needed the additional reset.

The sim is still using the wrong cores. Other games can see a 25% performance loss when using both CCDs and these hybrid CPUs are the present\future. This sim needs to know how to handle them correctly.

I suspect the sim is using the cores that clock higher, since the non-cache\non-gaming CCD cores clock higher. We could try downclocking the cores on the second CCD to make them slower than the the gaming CCD cores. This way the fastest cores will also be the cached gaming-cores. Maybe DCS will handle it better and the devs could use this info to fix the programing? 

[Coug4r]

I'm on a 5800X and the first thing i also noticed was the fps fluctuation.

[Beardy McWhisky]

Hmm cant say its an issue Im having tbh.

5800X3D, 64GB DDR4, 4090

[elekrytyk]

The same fluctuations and CPU load.
7950X3D+4090
 

[Moezilla]

11 hours ago, Beardy McWhisky said:

Hmm cant say its an issue Im having tbh.

5800X3D, 64GB DDR4, 4090

The 5800X3D has a single CCD, so your data supports OP's contention that, on multi-CCD CPUs, DCS performance is degraded if both CCDs are used by the game.

[ColinM9991]

On 10/21/2023 at 2:29 PM, Luka CMF said:

like the code is purposely written in a way that it only fully utilizes intel CPUs.  

The multithreading FAQ does specifically reference P and E cores, so it certainly doesn't do itself any favours

[speed-of-heat]

On 10/21/2023 at 2:29 PM, Luka CMF said:

like the code is purposely written in a way that it only fully utilizes intel CPUs.  

nope, on my 12700k DCS uses 2 p cores and all of my e cores... 

Edited October 30, 2023 by speed-of-heat

[hazzer]

On 10/23/2023 at 6:53 PM, ColinM9991 said:

The multithreading FAQ does specifically reference P and E cores, so it certainly doesn't do itself any favours

Having just purchases a 14700k, I can tell you that it isnt much better.

[Obic]

I also have a 5800x3D + 4090 + 64GB ram & it’s been a cpu <profanity>show on my end too since 2.9 dropped.

vR is frame drop city as 2.8 was just silk smooth 80fps 

I’ll try that no shadow for fun when I fly back from work 

[Stewart]

Might want to try this guys and see if it works. 

 

 

[Deanzsyclone]

On 10/22/2023 at 10:25 AM, Spectre1986 said:

Thanks for the input Deanzsyclone, we have the same config, but my issue seems to start in the main menu and persists thru to in-sim at any settings. Can you load the sim up to the main menu, press RCtrl+Pause to show the frame time graph and get a screenshot ("prt sc" on keyboard) and paste (Ctrl+V) it here? I'd like to see if you have the same fluctuations. I'll also test your shadow settings recommendation and report back.

I just went and did the latest BIOS update for my "Asus Tuf Gaming x670e-plus WiFi" motherboard and then did all available driver updates, including GPU. It made no change for me. It's still loading the wrong cores and giving a terrible experience. I have included the screenshot from the "Huey Hard Ground Attack" quickstart mission on the default map. Note: I'm not loading any extra software (SimShaker, SimAppPro, etc.) for these tests, I'm only running Pimax and DCS.
 

Here is my newest dcs.log file with the expanded logging mentioned previously.

null

Sorry just seeing this now I don't really get any issues like that in the home screen. And now my VR headset has been returned to the manufacturer trying and waiting for the replacement one to arrive so I can't really do much at the moment.

[Archaic]

I have 7950X3d as well and now noticing stuttering even at main menu every 5-8 seconds. Didn't happen before this last update.

[Spectre1986]

I did some more testing using Process Lasso.

First, I used it to always limit DCS to the first 8 physical cached cores. So cores 0, 2, 4, 6, 8, 10, 12, and 14. I am excluding the entire non-cached CCD and also excluding virtual cores (SMT\HyperThreading) from the first CCD.

This provided the best experience yet, nearly locked at 90 FPS, with very short and rare dips down to 45 FPS. The experience was noticeably better as it's mostly held back by the FPS limit I've set. This testing is in the Huey Hard Quick Start. After I've killed everything, I land facing the column of burning and exploding units, then take the screenshot. I'm in the Pimax Crystal and for software running I have Pimax, Simshaker for Aviators, Simshaker Sound Module, Process Lasso, Discord, and Simapppro (along with DCS). These apps do not have affinity set, only DCS. I'm able to verify with Task Manager that DCS is only loading the physical cores on the cached\gaming CCD, so PL is working.

 

Here are the results without any affinity set in Process Lasso, this is just regular DCS using the standard code. It's the same mission, same settings, and a screenshot of the same burning units, but with nothing currently exploding (so it should be even easier to render). You can see the performance is not only cut in half, the frame time has much more "noise" when holding a constant FPS, the graph is rougher and the experience is VERY noticeably worse. Task manager verifies DCS is mostly using the wrong CCD, it's maybe using ONE core on the correct CCD, then dumping all the data across the slow Infinity Fabric...

 The DCS developers have failed to address this issue despite multitudes of feedback, expanded logging testing, and log delivery. Even though EVERY modern gaming CPU has a hybrid design (Intel has P and E cores, AMD has gaming vs productivity CCDs) they're stuck in the past and appear to just randomly assign threads to nonsensical physical and virtual cores. Luckily we're able to take things into our own hands and force it to be done correctly with Process Lasso. I'm using the free version but I'm not sure if there are limitations that'll force me to buy it.

[Aries]

Thanks on this info... I just upgraded to 7800x3D & 7900XTX and was having a lot of questionmarks cause this "beast" was behaving like the old 10600K/3090Ti next to it...

"Screenshot" is from 2D usage. not VR...

 

Edited November 19, 2023 by Aries

[Spectre1986]

On 11/19/2023 at 3:57 PM, Aries said:

Thanks on this info... I just upgraded to 7800x3D & 7900XTX and was having a lot of questionmarks cause this "beast" was behaving like the old 10600K/3090Ti next to it...

"Screenshot" is from 2D usage. not VR...

 

 

Your 7800x3D doesn't have a second, non-cached CCD. All of your cores are on the same CCD with the huge cache so your CPU should be performing near-peak, you could consider using Process Lasso to contain it to the even-numbered physical cores for a potential boost on the CPU side, but you're GPU limited according to the DCS frame counter in your screenshot.

That 7900XT isn't much better than the 3090ti, especially in VR, and DCS is an odd one that may prefer Nvidia and perhaps charges an AMD tax on the GPU side. It may be worth swapping GPUs and testing again.

Edited November 29, 2023 by Spectre1986

[MustangSally]

Has this been reported @NineLine @BIGNEWY I'm experiencing exactly the same......7950x3d with nVidia 4090??

[waistcat]

Has this issue been progressed any? I'm a 7950X3D & 4090 user and now have alopecia from all the pulling out my hair trying to get DCS to run without stutters. Process Lasso was as much good as a fart in a spacesuit to me but then again I don't know what the process cores are with this CPU and I'm not a CPU savvy as most of the clever people on this thread.

[speed-of-heat]

Have you tried this https://www.hardwaretimes.com/amd-enable-the-xbox-game-bar-on-the-ryzen-9-7900x3d-7950x3d-processors-for-better-performance/