RTX 4090 performance with VR (Ryzen 5900X / Reverb G2)

I don’t have a 4090, just a 3080ti, but I also am trying to run down where the stutters are coming from - theses are big pulses in the frametime and they happen about 3 or 4 seconds apart. I think they might be from Windows 11 22H2 but they are just more noticeable because we are getting very smooth otherwise with or without motion reprojection (MR). I see that my frametime is fairly constant (moving average is 23 msec or less which qualifies for 36FPS MR) but these “pulses” are easily visible. There were threads about “HolographicShell” removal with WMR but I don’t have that. Just looking for others - especially with 4090s - that should have conquered the performance issues but still have these giant stutters/studders/judders/pulses.

2 Likes

A-a-a-n-d I’m back to 20 fps. Tried to see if I can reduce scenery stutters, uninstalled couple of last Windows updates, changed NDU registry keys, killed Holographic Shell process. Cleared D3D cache, MSFS scenery indexes, deleted content.xml. And got this mess again. Reintalled updates, changed everything back one by one, but still stuck at 22 fps. Tried all kinds of changes, but it’s stuck there with CPU frametime of 40ms! So weird…

Wow, after nothing at all helped,I went into thte MSFS settings and turned ott AI Traffic. And got 30 fps instantly. What’s curious, enabling it again was fine. As if it was stuck at 20 fps until disabled and enabled again…

Edit: yes, I did some experimenting and traffic setting has a massive effect on my fps. Off immediately removes all problems. Live traffic usually has a small impact. AI with low settings can be OK, but generally AI traffic potentially just massively kills my fps. I think that was it all the time, because what I saw npw when I paid attention is that KJFK was just full of the same livery airliners, like maybe 40 of them were visible, when my AI settings were ar modest 40% ground aircraft. Sometimes MSFS was just dumping dozens of the same aircrafe there for some reason. And that was only the visible ones. So any of you - in case of any FPS problems - try turning off the traffic first and see if it’s the culprit. I now recall that I read that live traffic was an fps killer and turned on AI traffic for testing, reasoning that it would be more stable and predictable and is better for testing. So many lost hours gone to waste…

I also tested OpenXR resolution at 150% and 125% as someone had suggested. Surprisingly it wasn’t affecting fps as much as I thought. But I didn’t see any stability improvement either. I still have a compbination of extremely smooth visuals with 30/45 fps MR, but frequent scenery stutters every few seconds. It’s not bad, I guess I can live with that, but I wish I could get rid of the stutters…

Considering you have some of the best hardware available right now, the existence of those stutters is really disheartening. Maybe you could make a dedicated VR thread for the stutter issue, and people could report whether they get them or not, and we could collect data and try to figure out what the common root cause is. Considering some people don’t seem to get them.

You’ve got no BestBuy, Walmart, B&H, … nothing near you? Even so, according to my plan, you only have to wait for one day. The next day, scalpers are not interested anymore because those willing to pay $3000 will have gotten one already

Because they might be interested in building a reputation and brand loyalty. If the goal is to make as much money as possible, they would have set the price at $3000 and then gradually reduce it over time (essentially become scalpers themselves).

Yes, sometimes, and only if I concentrate and look for it. My guess is that the MR is switching between different reprojection rates.

I thought so too, but it happens also in a very stable flight with a really stable and locked 30 fps MR, and when it’s locked at 30 and has 35% overhead it’s not going to jumpe to 22.5 and back. Still, if I look sideways - there are those scenery stutters every few seconds. Not exactly same amount of seconds, they do feel random - sometimes 3 seconds can pass between stutters, sometimes 15 seconds. If I just look ahead I may not notice them at all. But sightseeing and looking left they are very noticeable.

Well, it’s going to be another stuter thread, but I guess why not? It is a separate topic, so yes, I’ll do that…

Remember, in Open XR toolkit you can locked FPS for VR too :slight_smile: if they jump too often.

Ohvok, didnt realize there was already a VR stutter thread. Is it still alive?

I’m sure there are several, but I started my own: https://forums.flightsimulator.com/t/troubleshooting-scenery-stutters-on-high-end-hardware-rtx-4090-3900x-reverb-g2/

I do that, but it doesn’t really help. According to the developer, latest OpenXR versions don’t need locking as much as the transitions are smoother.

so im generally cpu limited with my 12900k and 3090 - i think primarily because of LOD/3rd party airports and vatsim traffic. airborne FPS are generally pretty good (40-50 but on ground im usually 20-30).

given that i am cpu bound - will a 4090 bring about zero improvement in performance? i understand i can probably bump up settings and probably get better quality but does being CPU bound mean that your FPS are maxed out unless the CPU is improved?

why do you think you are cpu limited?
what does the developer mode fps counter tell you (limited by main thread or limited by gpu)?
do you use open xr toolkit? if so, using the advanced fps counter what numbers do you get for cpu and gpu?
task manager can be useful, take off or peek out the bottom of your headset and run it, select performance tab, what % utilization is your cpu and gpu hovering around?

The CPU frametime reading is not valid in the OXRTK, or rather it’s incomplete. Enable the overlay in OpenXR Developer Tool, that will tell you proper frametimes.

I think that with hi-res VR we should all be GPU-limited in native resolution, but results can vary with MR and DLSS etc.

We seem to have been following a parallel path the last few years from the early days of CptLucky’s 2070s guidelines. My upgrade path has been:
3950x+2070s
3950x+3080
5900x+3080
5900x+3080ti
5800x3d+3080ti
5800x3d+3090ti

Whilst all the GPU upgrades came with FPS increases (which I then promptly got rid of by bumping up settings :slight_smile: ), the biggest reduction in stuttering came with the 5900x->5800x3d. It really does feel like with MSFS ‘cache is king’!

That said, I don’t use MR. I didn’t like the shimmering that came with it and am not particularly sensitive to low FPS in VR. Plus, I moved to a Varjo Aero a few month back so MR is not an option.

I picked up a 4090 today (same Gigabyte Gaming OC as you) and have since learned that the 4090 does not play nice with the Aero so I may dust off the old G2 and run a few benchmarks. The only significant difference between our systems is the CPU so it will be interesting to see how the results compare.

Absolutely - don’t forget to share how it goes. As I mentioned, I am planning to upgrade when the next X3D CPU is out. Looks like January-February if we’re lucky. 5800X3D would be a downgrade for me in terms of work from 5900X’s 12 cores, as I use this PC as a main workstation. (ona separate boot/partition/Windows from MSFS). 7900X3D shoud be a good upgrade though. If you can call a whole new system an upgrade, as it’s a new AM5 socket, so new motherboard with a new DDR5 RAM.

My cpu core count has been going in the opposite direction to my cuda core count :slight_smile:

Zen4 x3d cpus will possibly be my next upgrade too - I’ll see how the 4090 works out with the Aero once they’ve fixed the issues. With the 3090ti I am already getting consistent 40-50fps with the Aero bumped up to 37ppd and most settings on high/ultra (using dynamic foveated rendering helps). I really don’t need any more than that to be honest and could stop there and be happy but these incremental improvements are quite addictive so I’ll probably not stop until everything is maxed out at 39ppd and I’m still getting 50fps !

2 Likes

Now that’s what I’m talking about! Thanks for that link, will order one when it comes out. :grinning:

So I have dusted off the G2 and done a bit of benchmarking. Nothing scientific - just eyeballing the frametimes and indicated FPS during a couple of repeatable scenarios. I haven’t performed any testing with MR enabled yet since I tend not to use it.

My specs are Gigabyte Aorus x570 Master, AMD 5800X3D, 64GB of G.Skill CL16 3600MT/s RAM and an Asus ROG Strix water cooled 3090ti (replacing with a Gigabyte Gaming OC 4090 - only model in stock in NZ). Nothing overclocked.

My settings in MSFS are:

TAA 100% render scale
OpenXR 100% render scale
OpenXR Toolkit FFR using the quality/wide preset

All MSFS VR settings on High except:
Contact shadows and light shafts low
Ambient occlusion, raymarched reflections and bloom off
TLOD 200
OLOD 150
Anisotropic filtering 16x
Texture supersampling 4x4
Shadow maps 1024
Terrain shadows 256
Cubemap reflections 64

I ran 4 tests:

  1. My standard test of the Baron G58 at La Guardia on the scattered cloud preset taking off and flying to Manhattan
  2. The Girona landing challenge in the TBM
  3. The Jackson landing challenge in the CJ4
  4. Your test of the PMDG 737 taking off from JFK in overcast weather

First the 3090ti results:

  1. 40FPS on the runway dropping to 36FPS over NYC. Noticeable scenery micro-stuttering looking out to the left. Dropping TLOD to 100 gave me 43FPS on the runway and 40-43 over NYC
  2. 39-44FPS. Dropping TLOD to 100 made no real difference - maybe a more consistent 42-43FPS
  3. Rock solid 46-47FPS all the way down. Didn’t try TLOD 100.
  4. 40FPS on the runway. Dropped as low as 29 flying low over Manhattan.

The 4090 results:

  1. With TLOD200 I had 40FPS on the runway again (no improvement). This improved to 48FPS approaching Manhattan but then dropped to 30-35FPS flying low between the building of downtown Manhattan. Interestingly the OpenXRTK overlay was not indicating that I was CPU bound but, like you, when I opened the OpenXR overlay there was a 10ms discrepancty and I was clearly CPU bound. Looking at the Windows perf monitor, the 4090 was virtually idling at 4%!! The possiblity for upping the GPU intensive settings would appear to be huge :slight_smile: The other big difference is that the micro-stuttering is completely gone. Dropping TLOD100 confirmed the CPU bottleneck with me now seeing 55FPS on the runway with it never dropping below 52FPS even when snaking between the downtown buildings. This is simply incredible - completely butter smooth, full resolution and zero stutters maintaining over 50FPS in one of the heaviest scenery areas in the game. I checked the perf monitor - GPU at 11%!!

  2. Similar improvements on the Girona landing challenge. TLOD200 gave me 52-54FPS whilst TLOD100 was 61-62FPS

  3. The Jackson landing challenge is the least demanding of these 4. I didn;t bother with the TLOD100 test but TLOD200 was giving me 62FPS already all the way down with the GPU at 15% (I have seen the GPU hit more than that just sitting at the MSFS menu!)

  4. I didn’t see any improvement in FPS with the 737 at TLOD200 over the 3090ti results apart from maybe higher lows over downtown. I didn’t check frametimes but I suspect that I am CPU bound from the outset (given that the GPU was sitting at 21%). The big difference though was that without any micro-stutters it just felt far, far smoother.

I tried SU11. No huge improvement for me. Maybe a few extras FPS over the heavy scenery areas. I think I was seeing a fairly solid 41-42 over downtown NYC in the Baron (vs ~35). Other noticeable difference was that the GPU was now idling at 0% in the menu rather than 10-15% (in low power mode). Too many of the planes I fly don’t currently play nice with SU11 so went back to SU10. Unfortunately that process is a bit flawed and I am now half way through reinstalling :frowning:

Bottom line is that even with a 5800X3D I get CPU bound pretty quickly with some settings but the scope for increasing the non-CPU intensive settings appears to be huge. Once the 7900X3D arrives I might finally reach my full resolution Aero, everything ultra, 50FPS nirvana :slight_smile:

5 Likes

Thanks for the detailed post. I wish the TLOD could be optimised or shifted to the GPU somehow if there are such bottlenecks even with 5800x3d