Z's 4090 Thread

[LazyGamer]

What are you using to up the voltage? My slider is greyed out in Afterburner.

You may have to unlock it in the settings. Check there?

 

[Zarathustra]

You may have to unlock it in the settings. Check there?

Ah. Yes. Now that you say it, I vaguely remember this being an option. I probably did it when I first got my Pascal Titan years ago and then forgot about it.

Problem is it is difficult to correlate the resulting slider that goes from 0 to 100 with real voltage levels, and I have no idea how much is too much.

Unlike some previous cards for which the voltage slider was just there for placebo, this one appears to actually do something, but it doesn't appear to do much, as the overall power output doesn't change a whole lot.

So just a heads up that voltage adjustment can absolutely have a positive effect on the 4090, and 3090 Ti cards. I've been experimenting I'm just under 3 GHz, stable, for gaming on the 4090 (I was able to go over 3 GHz for some benching but it wasn't stable for gaming) and matching what Brent achieved in his 3090 Ti review at 2160-2175 Mhz. Both cards are using around ~450 watts and thanks to the hybrid liquid cooling solution they have they're holding 60-64c at stock fan settings. There are moments when they use less and more power but mostly I saw them hanging around 430-450w. This time around testing was done while playing The Witcher 3 and Hogwarts, and then the canned benches for CB2077/Metro Exodus.

As always use caution when adjusting voltage but at least for me it was the only way to get to the higher clocks w/o crashing. It was probably the silicon lottery that prevented me from just upping the clocks and power limit to get there. Air cooled cards can probably do about the same at the expense of cranking the fans a bit higher but thanks to the huge heatsinks these cards often have I could see it happening. I've got a 3090 Ti identical to the one that @Brent_Justice used in his review in the 4930K rig and I'll do some more testing with it next weekend.

I was able to get mine stable at 3105Mhz in Starfield. It was a bit tricky to get a stable configuration, as the top clock seems to change both with the clock speed adjustment AND with the voltage slider.

I noted that it would always crash the game after a while if it was allowed to reach 3120Mhz, even with the voltage up at +95, so I instead started backing off the clock speed and increasing the voltage slightly to reach stability. Eventually I got stable at like voltage +97 and clock +235, but this is with it running at 39C-41C at full load. If allowed to run hotter, it would likely not be happy at those settings.

I have not played around with RAM clocks yet. It is tricky to do with Starfield, as it sometimes has what looks like RAM artifacts (sparkles and stuff) but aren't actually RAM artefacts, and this has only been made worse since I applied the reshade mods for HDR.

Back in the day we used to do these overclocks in a worst case benchmark (furmark or something) to stress test them, knowing that if they were stable there, they were stable in everything. These days you have to seemingly tweak them on a per-title basis to reach max performance with stability, which is a little annoying.

 

[Peter_Brosdahl]

These days you have to seemingly tweak them on a per-title basis to reach max performance with stability, which is a little annoying.

Yep, so true. I've tweaked mine higher but usually just leave in the 2895-2910 range with mem either at stock or 11,000. Seems to be the most compatible across a number of different engines. The few crashes I still get tend to be for the game itself as when I set everything back to default the same crashes still happen.

 

[Zarathustra]

Yep, so true. I've tweaked mine higher but usually just leave in the 2895-2910 range with mem either at stock or 11,000. Seems to be the most compatible across a number of different engines. The few crashes I still get tend to be for the game itself as when I set everything back to default the same crashes still happen.

I did some more playing around last night. a max boost clock of 3105 is still stable for me. I played with upping the ram clock. Some were talking about being able to run +1300 on the RAM, so I started with +500, no problems at all, and then tried +1000 which resulted in an insta-crash +750 results in rare crashes, +725 in even more rare crashes. I dropped it a little further, can't remember if it was all the way down to +700, but it has been stable ever since.

So my top settings appear to be:
Core 3105Mhz
RAM: ~11220Mhz.

Compared to the 2805Mhz I was consistently getting at default settings before tweaking (which is much higher than the 2520Mhz OC boost clock printed on the box), that is an approximate 11% gain, which I guess is nothing to scoff at by modern standards.

Gone are the days of crazy high overclocks. My GTX 470 back in in 2010 was in an air starved Shuttle SFF case still managed a ~35% overclock with a factory blower cooler. I mean, it got hot. The sticker with an artistic rendering of a girl with bikini armor and a sword on the side of the thing was slowly melting and sliding off the side of it, but it got me through countless hours of S.T.A.L.K.E.R. at 2560x1600

Heck, my 486 sx25 ran at 50Mhz no problem after I added an HSF (it shipped without any cooling at all) My Duron 650Mhz back in college ran at 950Mhz (Toi my great disappointment I never got it stable at themagic gigahertz mark)

Modern overclocks are usually pretty marginal in nature compared to the good old days, but the 4090 is bucking the trend, at least a little.

 

[Peter_Brosdahl]

So my top settings appear to be:
Core 3105Mhz
RAM: ~11220Mhz.

Well, you've inspired me. I'll try some more tinkering on the weekend but mine was really finicky when it got above 3GHz. As far as memory goes I had the exact same experiences as you but I've recently just been leaving it at +500 but too remember being able to do around + 700-750 on it. Have you increased the voltage on it? I think I'm running at 1.7 or so on mine. I tried stock but it really put up a fight then.

 

[Zarathustra]

Well, you've inspired me. I'll try some more tinkering on the weekend but mine was really finicky when it got above 3GHz. As far as memory goes I had the exact same experiences as you but I've recently just been leaving it at +500 but too remember being able to do around + 700-750 on it. Have you increased the voltage on it? I think I'm running at 1.7 or so on mine. I tried stock but it really put up a fight then.

I don't know what exactly the voltage setting means. In order to get 3105 stable I had to crank up the voltage slider to +98.

It's challenging because the frequency and voltag4e sliders are not completely independent of each other.

Raising the frequency slider will allow the card to boost higher (and possibly crash if the voltage is too low) but raising the voltage slider will also tell whatever algorithm they use that there is more headroom ans also cause it to boost higher, and potentially crash.

Once I established that 3105 was probably the ceiling, and 3120 was not stable, I tried dropping the frequency slider a little while raising the boltage slider a little, but every time I raised the voltage slider, it tried to boost even higher

In order to have the desired effect (keep 3105 as the max, but raise the voltage) I had to simultaneously increase voltage and drop frequency, and by a lot too. I had originally been up at ~+450 Mhz when I discovered that 3105 was probably going to be the max, but gradually as I rased the voltage that had to come down to about 332Mhz while raising the voltage for it to be stable and not try to jump up to 3120 Mhz

As far as the voltage goes, I now have a Afterburner preset saved at +98 with these settings. What does +98 mean? Who the hell knows. I don't have a real world voltage readout anywhere. Maybe if I open up GPU-Z or HWMonitor or something, but I was tired last night and didn't think of that.

I may have been aided significantly by the fact that I have been able to maintain a 100% GPU load temp of ~40°C ±1°C, a benefit of running my desktop in the tiny air conditioned server room, and being able to run fans at 100% because who cares when you can't hear it, right?

I don't think it raises the voltage by a lot though, as the power reading at max load only increased marginally was doing all of this, and still fit well within the 450W+ 6% cap of my card. I hope +98 is not dangerous for the card, but I think it isn't. Nvidia seems to have been pretty conservative when it comes to what they will allow here. I don't think we'll see damaging voltages without doing something drastic, like a hardware shunt mod.

I could be wrong though. This evening I should look at a different tool and see if I can read what real world voltages I am actually hitting.

 

[Peter_Brosdahl]

As far as the voltage goes, I now have a Afterburner preset saved at +98 with these settings. What does +98 mean? Who the hell knows. I don't have a real world voltage readout anywhere. Maybe if I open up GPU-Z or HWMonitor or something, but I was tired last night and didn't think of that.

I'm pretty sure I've got mine at around + 40-50. Be careful as I've heard this setting can cause a card, or more specifically the actual GPU to burn itself out. You can set the AB overlay to display it and I've heard that around 1.9V is the maximum recommended, but can still be risky. I've tinkered in the 1.7 to 1.9 range with my settings but less might be better. I would always defer to what @David_Schroth and @Brent_Justice say on this but that is what I've heard.

p.s. Yep, I too was experimenting with this setting in trying to break that 3 GHz barrier. I just got nervous about it after reading some cautionary tales from others on the internet.

Edit: Just to add that if you're keeping this thing at 40-50c then the above may not matter. I'm averaging between 55-65c with fans on default.

 

[Zarathustra]

I'm pretty sure I've got mine at around + 40-50. Be careful as I've heard this setting can cause a card, or more specifically the actual GPU to burn itself out. You can set the AB overlay to display it and I've heard that around 1.9V is the maximum recommended, but can still be risky. I've tinkered in the 1.7 to 1.9 range with my settings but less might be better. I would always defer to what @David_Schroth and @Brent_Justice say on this but that is what I've heard.

p.s. Yep, I too was experimenting with this setting in trying to break that 3 GHz barrier. I just got nervous about it after reading some cautionary tales from others on the internet.

Edit: Just to add that if you're keeping this thing at 40-50c then the above may not matter. I'm averaging between 55-65c with fans on default.

I loaded up HWmonitor during my testing.

With the voltage slider at +0, it was reading 1.050v. The highest it went at +100 was 1.100v

I have previously heard that Nvidia likes to cap their recent GPU's at 1.1v, so this makes sense to me.

I did suffer a little bit of instability with a crash after about 2 hours in game today, so I had to back off a little bit from 3105Mhz down to 3090Mhz

 

[Peter_Brosdahl]

With the voltage slider at +0, it was reading 1.050v. The highest it went at +100 was 1.100v

I have previously heard that Nvidia likes to cap their recent GPU's at 1.1v, so this makes sense to me.

Yep, I missed I zero in there so 1.090 to, as you mentioned, 1.1. I was at work and going off of my own memory. I've managed 2895-2910 (2925 when I can keep it under 55c) @ 1.070, here's a screenshot of my settings. This is near 100% stable for most game engines and a lot of the time when something does crash I'll try it again with default settings and still experience the same crash so I tend to blame the game at that point. It's pretty rare the opposite is true. I might tinker with it a bit more this weekend after hearing of your experiences.

 

[Peter_Brosdahl]

So I spent a good chunk of today tinkering with it and was able to consistently break the 3 GHz barrier but it took a big hammer. Basically had to crank the voltage to +100%, pushed the core clock to around +165, and tinkered with mem between 750-1450. I never saw your speeds but did manage to have it hang between 3000-3015. The power draw was between 480W and 520W at that point. I kept all of my fans at 100% and it was hanging around 58-59C after about an hour of testing. Good for bragging rights but not optimal for gaming as my goal these days is the highest amount of performance with the least amount of noise. I tried pushing it further but it just got too crash happy.

I tested with the following games as all have RT (and I copied the latest DLSS dll files into the older games). I picked these games because, besides being among my favorites, they all use different engines. All at 4K. CB2077 was used the most due to all of its RT features.

Cyberpunk 2077
The Witcher 3 Enhanced (DX12)
Metro Exodus
Shadow of the Tomb Raider (outdated but one of the few games I can turn DLSS off and then max RT w/ SMAAx4)

David has reminded me on a number of occasions that when overclocking VRAM that while it'll report successful higher frequencies it won't necessarily tell you about all the errors it's experienced at those higher speeds. Additionally, after a certain point, it's creating unnecessary heat plus taking power away from the GPU due to the actual board limit. That being said I backed it off a bit and found a new sweet spot which allows slightly higher speeds w/ default fan settings (which usually hang in the 40-50% range), about the same amount of heat and power draw as my previous settings, while squeezing a bit more out in performance.



CP2077, w/ RT settings maxed and DLSS qualtiy, pyscho settings but post junk off, averaged 86 FPS with this, 60-66c temps, 460-485W. Core clock hung around 2955-2970, and voltage 1.085-1.090. Still need to play the new DLC and looking forward to it with these settings.

Metro Exodus reported nearly identical FPS but I noticed the dips were not nearly as low as the previous settings, sometimes a difference of 10 FPS.

Witcher 3. Everyone nows that the DX12 mode is brutal and w/ RT features on max it can easily bring a 4090 to under 50 FPS at its lowest points. I saw 55-62 FPS for the lows and upwards of 80+ while wandering outside the Nilfgardian camp in Velen (find the missing soldiers mission).

SOTTR. Pretty much a breeze for a 4090 but nice to play a game w/ RT stuff maxed and not use DLSS. Basically anywhere from 80-120+ FPS.

Thanks for the inspiration, I could proably dial it in a bit more but these new settings allow the best of both worlds and even though I might not be pushing beyond 3 GHz I'll still grin as I look over at the rig and enjoy its near silent operation while using them and still be within 1-3% of those higher speeds.

 

[Zarathustra]

So I spent a good chunk of today tinkering with it and was able to consistently break the 3 GHz barrier but it took a big hammer. Basically had to crank the voltage to +100%, pushed the core clock to around +165, and tinkered with mem between 750-1450. I never saw your speeds but did manage to have it hang between 3000-3015. The power draw was between 480W and 520W at that point. I kept all of my fans at 100% and it was hanging around 58-59C after about an hour of testing. Good for bragging rights but not optimal for gaming as my goal these days is the highest amount of performance with the least amount of noise. I tried pushing it further but it just got too crash happy.

I tested with the following games as all have RT (and I copied the latest DLSS dll files into the older games). I picked these games because, besides being among my favorites, they all use different engines. All at 4K. CB2077 was used the most due to all of its RT features.

Cyberpunk 2077
The Witcher 3 Enhanced (DX12)
Metro Exodus
Shadow of the Tomb Raider (outdated but one of the few games I can turn DLSS off and then max RT w/ SMAAx4)

David has reminded me on a number of occasions that when overclocking VRAM that while it'll report successful higher frequencies it won't necessarily tell you about all the errors it's experienced at those higher speeds. Additionally, after a certain point, it's creating unnecessary heat plus taking power away from the GPU due to the actual board limit. That being said I backed it off a bit and found a new sweet spot which allows slightly higher speeds w/ default fan settings (which usually hang in the 40-50% range), about the same amount of heat and power draw as my previous settings, while squeezing a bit more out in performance.

View attachment 2746

CP2077, w/ RT settings maxed and DLSS qualtiy, pyscho settings but post junk off, averaged 86 FPS with this, 60-66c temps, 460-485W. Core clock hung around 2955-2970, and voltage 1.085-1.090. Still need to play the new DLC and looking forward to it with these settings.

Metro Exodus reported nearly identical FPS but I noticed the dips were not nearly as low as the previous settings, sometimes a difference of 10 FPS.

Witcher 3. Everyone nows that the DX12 mode is brutal and w/ RT features on max it can easily bring a 4090 to under 50 FPS at its lowest points. I saw 55-62 FPS for the lows and upwards of 80+ while wandering outside the Nilfgardian camp in Velen (find the missing soldiers mission).

SOTTR. Pretty much a breeze for a 4090 but nice to play a game w/ RT stuff maxed and not use DLSS. Basically anywhere from 80-120+ FPS.

Thanks for the inspiration, I could proably dial it in a bit more but these new settings allow the best of both worlds and even though I might not be pushing beyond 3 GHz I'll still grin as I look over at the rig and enjoy its near silent operation while using them and still be within 1-3% of those higher speeds.

I've had to gradually back off my max settings a little bit. I got very infrequent game crashes I couldn't' detect without being in game for several hours. So I ahve been stepping back bit by bit. I'm now down to a max clock of 3075 in Starfield, and thus far this looks like as far as I'm going to have to back off, but who knows.

 

[Peter_Brosdahl]

I've had to gradually back off my max settings a little bit. I got very infrequent game crashes I couldn't' detect without being in game for several hours. So I ahve been stepping back bit by bit. I'm now down to a max clock of 3075 in Starfield, and thus far this looks like as far as I'm going to have to back off, but who knows.

Yep, that's the way of it, having to play a game for hours for the true test. I had a crash with these settings while trying to play Marvel's GOTG the other night when I was barely 30 minutes in. Ironic since they worked after hours of "canned" benches with some actual game-play time.

 

[David_Schroth]

Stability will greatly depend on the game and even what areas of the game you're in...

 

[Zarathustra]

Stability will greatly depend on the game and even what areas of the game you're in...

I know.

I miss the days of picking an overkill looping benchmark, maximizing settings, and knowing it would be stable for everything.

Overclocking today is endless per title tweaking, and it is honestly rather annoying.

 

[LazyGamer]

Gaming has long been the 'ultimate benchmark'; stability issues not teased out by any number of synthetic workloads - many downright abusive - can surface just by having so many working parts interrelatedly stressing the system as a whole.

For my personal desktop (12700K, 3080 12GB, custom loop in O11 XL), I've pretty much stopped overclocking altogether. I'd previously gotten Windows to the point that it would no longer update to the next feature release and have just wiped and built it back up. A big part of that is most likely the memory overclocking I had attempted, which resulted in many a bluescreen.

This is not to say that I won't try again later, but for now, I'm enjoying having a working system!

 

[Brian_B]

Overclocking today is endless per title tweaking, and it is honestly rather annoying.

Yup, and for fairly minimal gain past what the hardware will already do automatically. It ain't what it used to be.

I still enjoy putting rigs together when I have the money to do it right, but overclocking, for me, is dead.

 

[Zarathustra]

For my personal desktop (12700K, 3080 12GB, custom loop in O11 XL), I've pretty much stopped overclocking altogether. I'd previously gotten Windows to the point that it would no longer update to the next feature release and have just wiped and built it back up. A big part of that is most likely the memory overclocking I had attempted, which resulted in many a bluescreen.

This is why I usually dump an image of my windows partition on a semi-regular basis, so I can revert if something goes wrong.

Having a huge NAS helps here. I can just boot up from a linux USB stick, and use dd (or the GUI disks app) to just dump an image over the network.