Dlss4 + 4090 = lower max oc than with dlss3

[yzonker]

Well you can debate all day long what OC is stable and which is not. Several of us over on the OCN forum noted that the new transformer model is more sensitive to core overclock. It's about 3 bins lower (45 mhz). Although Spiderman 2 with transformer may be even require one more bin lower yet.

The OC I was running always passed tests like Port Royal stress test, etc.. But it proved unstable in CP2077 with transformer active.

One thing I found is that looping R24 on the 4090 will find about the same max stable core clock as what I need for CP2077 now.

So unless you've run every game/test/etc... in existence and any that will exist in the future (impossible of course), you may always find something that isn't stable eventually.

 

[Dragam1337]

Well you can debate all day long what OC is stable and which is not. Several of us over on the OCN forum noted that the new transformer model is more sensitive to core overclock. It's about 3 bins lower (45 mhz). Although Spiderman 2 with transformer may be even require one more bin lower yet.

The OC I was running always passed tests like Port Royal stress test, etc.. But it proved unstable in CP2077 with transformer active.

One thing I found is that looping R24 on the 4090 will find about the same max stable core clock as what I need for CP2077 now.

So unless you've run every game/test/etc... in existence and any that will exist in the future (impossible of course), you may always find something that isn't stable eventually.

Very interesting - do you have a link to the convo ?

DLSS 4 crashes with the Transformer model is a widespread problem - just Google "Cyberpunk 2.21 crash" and you'll see many reports. It's also affecting other games that use the DLSS 4 Transformer model. The issue, as mentioned, is that DLSS 4's Transformer model is making stable UV or OC profiles unstable for many users. This is also true in my case with the RTX 4070 Ti, where I have a UV profile with an OC (1.04V +150 MHz) stable for last year in many games including Cyberpunk DLSS 3. It seems that the problem is related to ray reconstruction. If ray reconstruction is turned off, your game should not crash. You can also avoid crashes by using the older DLSS 3 ray reconstruction DLL in game (only the ray reconstruction DLL, other DLLs can be DLSS 4).

If you use the CNN model, crashes will happen much less frequently, but crashes are still possible, at least in my case. I solved it by lowering my OC by 15 MHz (to +135 MHz), which stabilized the CNN model and prevented crashes. However, the Transformer model still crashes roughly once every 30 minutes (but with +150mhz it crashes in 2-3 minutes).

I play at DLDSR resolution (which uses Tensor Cores, so it might be making things worse in my case), and there's only a minimal visual difference between Transformer and CNN model. So, I prefer to use the CNN model - it runs on lower power so the GPU is cooler and quieter. With the Transformer model, the GPU runs hotter and louder, and I have to find a less effective undervolt to keep it stable, which only makes it hotter and noisier. In my case, CNN model is about 15-20w less power consumption, around 1-2°C cooler, and roughly 150 fewer RPM on the fans (1900 vs 2050 rpm max, what really makes a difference to my ears ). So, until there is the option to switch to old model, I prefer using the CNN model.

Exactly what i was looking for people to share ! sounds like you experience the same thing as me, although to a lesser extent.

 

[yzonker]

Very interesting - do you have a link to the convo ?



Exactly what i was looking for people to share ! sounds like you experience the same thing as me, although to a lesser extent.

This was one of the latest comments. It had come up previously too though. Seems to be about 45mhz less than the previous release of CP2077 like I said though.

[Official] NVIDIA RTX 4090 Owner's Club

Last Updated: January 22, 2023 Note: This content is licensed under Creative Commons 3.0. This means that you are free to copy and redistribute this material, but only if the following criteria are met: 1) You must give appropriate credit by linking back to this thread. 2) You may not use this...

www.overclock.net

 

[Dragam1337]

This was one of the latest comments. It had come up previously too though. Seems to be about 45mhz less than the previous release of CP2077 like I said though.

[Official] NVIDIA RTX 4090 Owner's Club

Last Updated: January 22, 2023 Note: This content is licensed under Creative Commons 3.0. This means that you are free to copy and redistribute this material, but only if the following criteria are met: 1) You must give appropriate credit by linking back to this thread. 2) You may not use this...

www.overclock.net

Damn, some of them crashing with dlss4 with stock settings - saw steve from gamers nexus saying the same thing about the 4090 they are using for their benchmarking.

 

[oxrufiioxo]

For me in CP2077 I sit at the same 2955mhz on both models They both crash If I try to push beyond that I don't think it's a power issue because I was sitting at 460w regardless and can hit 575w in 3dmarks speedway.

I did not test a bunch of different areas just the market which seems to be one of the heavier areas or before the latest patch. I want to say it was 2980-2990 when path tracing first got released though so I am seeing some regression but on both models.

 

[Dragam1337]

For me in CP2077 I sit at the same 2955mhz on both models They both crash If I try to push beyond that I don't think it's a power issue because I was sitting at 460w regardless and can hit 575w in 3dmarks speedway.

I did not test a bunch of different areas just the market which seems to be one of the heavier areas or before the latest patch. I want to say it was 2980-2990 when path tracing first got released though so I am seeing some regression but on both models.

I'd say that you got a very good clocker on your hands then Ty for sharing !

To me this does suggest that individual chips very alot in regards to the "strength" of their tensor cores. Like @yzonker linked, some people are even crashing at stock settings, and have to lower clocks below stock with dlss4 :O

 

[oxrufiioxo]

I'd say that you got a very good clocker on your hands then Ty for sharing !

To me this does suggest that individual chips very alot in regards to the "strength" of their tensor cores. Like @yzonker linked, some people are even crashing at stock settings, and have to lower clocks below stock with dlss4 :O

It sit's at 2800 stock without me doing anything some of the others I worked with sit closer to 2600.

I don't think it's a crazy clocker though I think the best I managed was like 2995 and I've seem some over 3ghz. Not that it actually matters to performance lol.

 

[Dragam1337]

It sit's at 2800 stock without me doing anything some of the others I worked with sit closer to 2600.

Mine (despite being same model as yours) is at 2760-2745 stock, so yeah, deffo some chip to chip varience. It seems mine has strong raster cores, and mid tensor cores, with regards to how it behaves with dlss4.

To you edit : yeah, at 1.1v mine can do 3070 mhz (when not dlss4... lol) - it does boost performance a bit, but not as much as it boosts powerdraw... lol

 

[oxrufiioxo]

Mine (despite being same model as yours) is at 2760-2745 stock, so yeah, deffo some chip to chip varience. It seems mine has strong raster cores, and mid tensor cores, with regards to how it behaves with dlss4.

The tough thing with overclocking is there are too many factors.... I've even seen windows update causing shenanigans and if I run dual monitors it changes a tiny bit, and even driver to driver and game patch to game patch..... 330w power limit just always works lol.

I'm also not going to move my PC to my other room just to do this test on a native 4k display so I used DLDSR which could have also been a factor although I doubt it....


Although given that others are having the same issue I would say it's the switch to the Transformer model in this case.

 

[Dragam1337]

The tough thing with overclocking is there are too many factors.... I've even seen windows update causing shenanigans and if I run dual monitors it changes a tiny bit, and even driver to driver and game patch to game patch..... 330w power limit just always works lol.

I'm also not going to move my PC to my other room just to do this test on a native 4k display so I used DLDSR which could have also been a factor although I doubt it....


Although given that others are having the same issue I would say it's the switch to the Transformer model in this case.

Indeed. This is the highest i can do now with dlss4, but watt is getting out of hand with those settings. I usually prefer to have a solid 2700 @ 950mv undervolt, but will use overclock sometimes when a bit extra juice is needed.

Spoiler

I don't think there is any meaningful difference between dsr and having a screen at that res

Yep - transformer model using tensor cores to a much larger degree. I would be very curious to see if the stronger tensor cores on the 5000 series completely negates this.

 

[harm9963]

Well you can debate all day long what OC is stable and which is not. Several of us over on the OCN forum noted that the new transformer model is more sensitive to core overclock. It's about 3 bins lower (45 mhz). Although Spiderman 2 with transformer may be even require one more bin lower yet.

The OC I was running always passed tests like Port Royal stress test, etc.. But it proved unstable in CP2077 with transformer active.

One thing I found is that looping R24 on the 4090 will find about the same max stable core clock as what I need for CP2077 now.

So unless you've run every game/test/etc... in existence and any that will exist in the future (impossible of course), you may always find something that isn't stable eventually.

100% or 98% , Port Royal stress test show your results.

 

[yzonker]

100% or 98% , Port Royal stress test show your results.

Don't have one with SysInfo active.

 

[Zazigalka]

See if the new hotfix driver fixes ray reconstruction related crashes, seen people reporting cp2077/aw2 no longer crashing on it.

 

[Dragam1337]

See if the new hotfix driver fixes ray reconstruction related crashes, seen people reporting cp2077/aw2 no longer crashing on it.

I will test it out

 

[oxrufiioxo]

Seems 5080 owners are having all sorts of issues on it.

 

[xorbe]

DLSS 4 crashes with the Transformer model is a widespread problem - just Google "Cyberpunk 2.21 crash" and you'll see many reports. It's also affecting other games that use the DLSS 4 Transformer model. The issue, as mentioned, is that DLSS 4's Transformer model is making stable UV or OC profiles unstable for many users. This is also true in my case with the RTX 4070 Ti, where I have a UV profile with an OC (1.04V +150 MHz) stable for last year in many games including Cyberpunk DLSS 3. It seems that the problem is related to ray reconstruction.

No, the problem would be related to new stimulus tickling hardware paths that aren't meeting timing due to user UV / OC. Solution, more V or less OC. The problem is not ray reconstruction. If changing driver version causes crashes to go away, it's just happenstance. The primary issue is still lurking (unstable UV/OC).

 

[TheDigitalJedi]

Indeed. This is the highest i can do now with dlss4, but watt is getting out of hand with those settings. I usually prefer to have a solid 2700 @ 950mv undervolt, but will use overclock sometimes when a bit extra juice is needed.

Spoiler

I don't think there is any meaningful difference between dsr and having a screen at that res

Yep - transformer model using tensor cores to a much larger degree. I would be very curious to see if the stronger tensor cores on the 5000 series completely negates this.

I'm late to the conversation but wanted to share that this was witnessed today. I read your post and decided to give this a try with a friend's PC. This was tested with the 4090 Founder's edition. Higher overclocks are achievable with DLSS3. His normal and comfort zone is 2950-2955mhz. We enabled DLSS 4 and fired up Black Myth Wukong. The game automatically executed another shader compilation. The settings are 4k resolution, Ray Tracing Very High and all Cinematic options enabled with Balanced set at 65.

We then proceeded to play the game with DLSS4 enabled with the overclock of 2955mhz core and 1000+ on the memory. Things seemed fine for a few minutes until the game crashed. With DLSS3 these clocks were fine. DLSS4 however was not stable. We then lowered the clocks to 2910mhz core and 700+ on the memory. Everything played fine with DLSS4 and all is stable.

The same scenario happened with Cyberpunk. The game played beautifully with 2910mhz on the core and 700+ on the memory with DLSS4, max settings, Path Tracing and Framegen.

Here's a few pics from Black Myth Wukong and Cyberpunk with DLSS4. Visual bliss!!!! Forgive the saturation in these pics. We forgot about the HDR setting while snapping jpegs with the Steam Overlay.

 

[Dragam1337]

See if the new hotfix driver fixes ray reconstruction related crashes, seen people reporting cp2077/aw2 no longer crashing on it.

Hotfix driver made no changes for me - dlss4 rayreconstruction still needs more volt / lower clock than dlss3 reconstruction.

 

[Zazigalka]

Hotfix driver made no changes for me - dlss4 rayreconstruction still needs more volt / lower clock than dlss3 reconstruction.

At least it's worth the extra power

 

[zyarra]

I don't see anything like that.
My 4090 runs at 3150mhz(pretty much constant) regardless of DLSS version.

 

[squallypo]

So i tested a bit further, and it appears that it is not actually dlss4 causing the extra load / need for higher voltages or lower clocks, but rather using ray reconstruction when also using dlss4. This is also the case with hogwarts legacy after it got dlss4 - enable RR, and max stable oc is reduced to +100.

I saw steve from gamersnexus talking about them experiecing instability with dlss4 aswell, so likely same issue they are running into.

https://steamcommunity.com/app/1091500/discussions/7/603020878754680624/?tscn=1740226375 pretty much whats been going on in this thread