yes, you are then still power limited but at a higher clock since you are undervolted. Set power limit to where you want and as long as you have the curve set, the GPU will provide max performance for that level of power (which is a variable based on task/game/engine/res)I used the curve in Afterburner, so the higher boost was still available, just at lower voltageUsed GPU-Z to monitor, the voltage was lower, power was still pegged at 350w though.
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Rtx 3000 series undervolt discussion
- Thread starter mnewxcv
- Start date
Digital Viper-X-
[H]F Junkie
- Joined
- Dec 9, 2000
- Messages
- 15,004
So in this case, how come at lower voltage, we are not getting less wattage? is it just pulling more amps to compensate?
Because while it is pulling less watts, it is also doing higher clocks, which results in more work being done for the same power draw. That's why you want to set a lower power limit so you can get close to stock performance for less power, rather than more performance for same power. I'm going to revise the guide to make it a bit simpler, but as long as you have a curve set, regardless of power limit, you are running more efficient than stock. If you leave stock power limit, that equates to more work getting done rather than power savings.
Digital Viper-X-
[H]F Junkie
- Joined
- Dec 9, 2000
- Messages
- 15,004
but it's not clocking any higher, it's about the same clocks ( I didn't raise the clocks at all), so at the same clocks, with less volts.
Did you do your offset in a multiple of 15MHz? Sounds like whatever game you are running is just simply power limited regardless. I don't claim to be an expert, but I've spent a lot of time tweaking this stuff since I got my 3090. In my experience, even without a power limit adjustment, you should see a slight increase in clocks, unless you are in your highest boost table.
Digital Viper-X-
[H]F Junkie
- Joined
- Dec 9, 2000
- Messages
- 15,004
I'll post pics when I get on the rig again
D
Deleted member 143938
Guest
Does anyone else have their voltage spike above the voltage they set in Afterburner?
Digital Viper-X-
[H]F Junkie
- Joined
- Dec 9, 2000
- Messages
- 15,004
you are definitely power limited. How aggressive is your offset?
So I used a variety of methods to undervolt on my STRIX 3080. The OP, locking on VF points, the one with the manual flat line above a certain voltage (which I was more familiar with from my 2080Ti) and ran Timespy Extreme more times than I care to admit.
In the end I was finding things that would pass tiemspy but with noticeable glitches (which would persist even after returning to stock), and then trying them in game (AC Odyssey) only to have it crash. So I walked back a bit from timespying and now I just run AC Odyssey at 4K Ultra for a bit with a conservative undervolt.
WIth +105 to a cold-stock curve up to 981mV (and then flat from there, which changes frequency depending on what temp is), I had a sustained boost of 1995 in AC Odyssey at a cool 986mV and 66 degrees, with a custom fan curve that stayed under 55% up til 70 degrees. (55 is inaudible to me in my setup, but I do hear them a bit at 65).
I was pretty happy with that, and probably could try 120 or 135 but I would definitely try AC Odyssey gaming over TSE in the future. More fun.
I tried to reproduce that result in GPU Tweak II using a 2005 boost and an 88% power limit, but it didn’t work. Clocks were jumping all over between 1850 and 1935, and temps were at 75+ (in Control, not AC Oydssey, which I have to try... in case the RTX was the driving factor there).
In the end I was finding things that would pass tiemspy but with noticeable glitches (which would persist even after returning to stock), and then trying them in game (AC Odyssey) only to have it crash. So I walked back a bit from timespying and now I just run AC Odyssey at 4K Ultra for a bit with a conservative undervolt.
WIth +105 to a cold-stock curve up to 981mV (and then flat from there, which changes frequency depending on what temp is), I had a sustained boost of 1995 in AC Odyssey at a cool 986mV and 66 degrees, with a custom fan curve that stayed under 55% up til 70 degrees. (55 is inaudible to me in my setup, but I do hear them a bit at 65).
I was pretty happy with that, and probably could try 120 or 135 but I would definitely try AC Odyssey gaming over TSE in the future. More fun.
I tried to reproduce that result in GPU Tweak II using a 2005 boost and an 88% power limit, but it didn’t work. Clocks were jumping all over between 1850 and 1935, and temps were at 75+ (in Control, not AC Oydssey, which I have to try... in case the RTX was the driving factor there).
Attachments
This doesn't make any sense... I switched a 3070 Gaming X Trio for a 3070 SUPRIM, and the results are really weird. I get lower score with 2070MHz @ 981mV on the SUPRIM compared to the 2040MHz @ 981mV on the Gaming X Trio. I tryed to increase the voltage to 1000mV, and it's still slightly behind...
Now, I can understand the chip may not be as lucky, but how is it even possible that a 3070 performs worse than another one at lower clock? This really doesn't make any sense... Also this graphics card always boosts +15MHz above what I set, while the other one had the tendency to do the opposite and boost -15MHz what I set (not always like this one tho).
Now, I can understand the chip may not be as lucky, but how is it even possible that a 3070 performs worse than another one at lower clock? This really doesn't make any sense... Also this graphics card always boosts +15MHz above what I set, while the other one had the tendency to do the opposite and boost -15MHz what I set (not always like this one tho).
are you setting your overclock/offset in a multiple of 15?
Yes, it was +150 on the Gaming X Trio, and it's +240 on the SUPRIM. I can't wrap my head around it... Unless there are multiple versions of the same chip, like A and non-A chips for the 20 Series?
are you talking about time spy? If so, could you share the links to the results?
Does TimeSpy save locally to a file? If not, I have nothing aside from the Unigine Heaven scores.
P.S. Did one more run at 2055MHz @ 981mV and this time I got a better score than the Gaming X Trio at 2040MHz, finally. Probably I did too many runs in a row before, or I don't know honestly... This graphics card is impressive for sure, being able to run 30MHz higher with the same voltage... Even tho it starts to get artifacts with TimeSpy, so not totally stable, that's why I lowered it. I could also run 2040MHz @ 950mV without any issues, even though I wouldn't know if it's because of the better components or if it's a higher binned GPU.
Last edited:
Hey mnewxcv, thanks for the technique advice. I’ve been playing around with my Zotac Trinity RTX 3090 this week and I’ll try your technique when I get back to it next week.
One thing I wanted to mention, although it isn’t necessary for your technique, is that you can hold ctrl while moving a node to shift the curve while keeping it anchored on either the right or left depending on what node you move. I was using it to raise my target mV node while keeping the left side of the curve anchored.
Lastly, two questions:
1) what’s the best thing for testing stability? I.e. what’s the least stable game? I understand this is a difficult question, but I’m getting tired of running the TimeSpy benchmark and stress test just fine and then having AC Valhalla crash after two hours...
2) are we sure it matters when (I.e. what temp) you apply your VF Curve? Or is +150@950mV going to produce the same result regardless? Like, I get that the curve itself will look different depending on temps, but will +150@950MHz actually result in different frequencies once stabilized at load depending on when you applied that curve? Put another way, when you hit apply does the GPU BIOS store the +150 offset or the result frequency. If the latter, then I understand why you’d need to be careful when you apply your curve, but the former seems like how it should work (and would mean you don’t need to care about when you set your curve).
One thing I wanted to mention, although it isn’t necessary for your technique, is that you can hold ctrl while moving a node to shift the curve while keeping it anchored on either the right or left depending on what node you move. I was using it to raise my target mV node while keeping the left side of the curve anchored.
Lastly, two questions:
1) what’s the best thing for testing stability? I.e. what’s the least stable game? I understand this is a difficult question, but I’m getting tired of running the TimeSpy benchmark and stress test just fine and then having AC Valhalla crash after two hours...
2) are we sure it matters when (I.e. what temp) you apply your VF Curve? Or is +150@950mV going to produce the same result regardless? Like, I get that the curve itself will look different depending on temps, but will +150@950MHz actually result in different frequencies once stabilized at load depending on when you applied that curve? Put another way, when you hit apply does the GPU BIOS store the +150 offset or the result frequency. If the latter, then I understand why you’d need to be careful when you apply your curve, but the former seems like how it should work (and would mean you don’t need to care about when you set your curve).
how do you anchor the left or right side of the curve? I will add that to the OP when I can figure out how to do it.
For stability, I don't want to suggest anything that isn't free in the OP. Bright Memory Infinite Raytracing Benchmark is free on steam and seems to stress the card more, BUT if you are hitting a power limit and therefore NOT maxing out your clocks, you might have instability at those higher clocks that aren't reached during the benchmark.
As for the temperature when setting the curve, it 100% matters. If you set +150 950mV on a cold card, it will run at a higher clock than if you set that point when it was hot. It only sets the clock speed once, based on where the curve is. Cold curve is higher. It sets the offset based on the curve where it is in that moment, and then locks it until you change other settings.
It always anchors the left or right, which is anchored depends on what node you are moving: nodes on the right of center anchor the left side, nodes on the left of center anchor the right side. This makes it slightly awkward to hit a frequency target on a left-side node (e.g. 900mV) while keeping the left side anchored (since you have to be moving a right-side node).
I get that reasoning (might be worth adding a post-script about if you’re willing to spend money what apps are best suited), but I’m still curious about what people have found to be the quickest way to find a fully stable VF Curve. You mentioned Hitman? Or Hitman 2 or something?
One thing you might consider changing in your OP is to suggest recording what your temperature is at boot and using that as your target temp for VF Curve setting since that’s when MSI Afterburner will quickly launch to set its stuff and then quit in daily use. Or are you suggesting getting the card as cold as possible for a reason? Like my card will boot at like 38-40C w/ fan disabled, but I can manually set fan to 100% and get it down to low 30s.
Oh, that lets you rotate the curve. So if you want higher offsets at lower mV, you can rotate it to reduce the upper end. That could come in handy if you want a full curve (not an undervolt).
So what’s the actual impact of temperature on a given frequency voltage node? At higher temperatures is the frequency lower for a given voltage? Do we have any idea by how much? Does it change every degree or are there bands (e.g. under-40, 40-50, 50-60, 70+)?
EDIT: I've been playing with this a little today and I found that it's fairly sensitive and also doesn't track temperature exactly, it seems like there's some time-at-temperature, or perhaps average-temperature in some time window. I saw the same result from [23*C,30*C) (2010MHz max boost) and then at 30*C I saw it drop to 1995MHz, 32*C dropped to 1980MHz and a lower voltage, 34*C dropped to 1965MHz. This would indicate that you need to think about boot temperatures a lot when you apply your curve, e.g. if you restart with a bunch of residual heat in your GPU, afterburner will apply your curve to a hot GPU Boost curve, which to me looks like it can "crush" the high end of your curve.
This would all be so much easier if the offsets themselves were stored and applied instead of a specific curve. e.g. +150 @ 925mV would just fluctuate with GPU Boost such that it might run in the [1855-1905] range.
EDIT2: So here's what I'm going with currently after playing around with it a lot today, it's a curve created by rotating the default one with the left-side anchored until my 900mV node hits it's 1845MHz target, and then I eliminate the rest of the curve. It's been stable so far but haven't done much long-play-session gaming (which is where I've had issues in the past). I played AC Valhalla for an hour and it's just sits at like 250W power usage while pegged to 900mV @ 1845MHz, which leaves my GPU fans at a quiet 58ish% and temps in the high 60s. 3DMark benchmarks show that the GPU Boost algorithm is moving the frequency around a lot in the 1700s and 1800s, averaging like high 1700s (e.g. 1784). My benchmark scores are all up ~4% compared to stock.
EDIT3: I just booted my computer this morning and the card was at like 23*C so the top of that curve is now 1860MHz @ 900mv instead of the 1845MHz in that graph when the card was 27*C. Seems like it's probably a good idea to give yourself some headroom to account for situations like this (15-30MHz seems about right).
EDIT: I've been playing with this a little today and I found that it's fairly sensitive and also doesn't track temperature exactly, it seems like there's some time-at-temperature, or perhaps average-temperature in some time window. I saw the same result from [23*C,30*C) (2010MHz max boost) and then at 30*C I saw it drop to 1995MHz, 32*C dropped to 1980MHz and a lower voltage, 34*C dropped to 1965MHz. This would indicate that you need to think about boot temperatures a lot when you apply your curve, e.g. if you restart with a bunch of residual heat in your GPU, afterburner will apply your curve to a hot GPU Boost curve, which to me looks like it can "crush" the high end of your curve.
This would all be so much easier if the offsets themselves were stored and applied instead of a specific curve. e.g. +150 @ 925mV would just fluctuate with GPU Boost such that it might run in the [1855-1905] range.
EDIT2: So here's what I'm going with currently after playing around with it a lot today, it's a curve created by rotating the default one with the left-side anchored until my 900mV node hits it's 1845MHz target, and then I eliminate the rest of the curve. It's been stable so far but haven't done much long-play-session gaming (which is where I've had issues in the past). I played AC Valhalla for an hour and it's just sits at like 250W power usage while pegged to 900mV @ 1845MHz, which leaves my GPU fans at a quiet 58ish% and temps in the high 60s. 3DMark benchmarks show that the GPU Boost algorithm is moving the frequency around a lot in the 1700s and 1800s, averaging like high 1700s (e.g. 1784). My benchmark scores are all up ~4% compared to stock.
EDIT3: I just booted my computer this morning and the card was at like 23*C so the top of that curve is now 1860MHz @ 900mv instead of the 1845MHz in that graph when the card was 27*C. Seems like it's probably a good idea to give yourself some headroom to account for situations like this (15-30MHz seems about right).
Attachments
Last edited:
I really don’t think it matters when you set your curve in terms of temperature. If you set it cold and then check it when it’s hot, it’ll be the same as if you set it when it’s hot.
I think all you’re doing is setting an offset per voltage bin, e.g. you set +60 at 900mV and it’ll try to use 5 frequency bins higher at that voltage bin. The frequency bins will change based on thermals, but it’ll try to use 5 bins higher.
My anecdotal evidence for this is that you can set your curve while hot and if you subsequently cool the card and then run under load, you’ll see the frequency go higher than what you saw when you set the curve, e.g. maybe 900mV was 1830 when you set that as the limit of your curve, but then it’ll boost to 1875 at 900mV after you let it cool and re-load the GPU without touching the curve.
I think all you’re doing is setting an offset per voltage bin, e.g. you set +60 at 900mV and it’ll try to use 5 frequency bins higher at that voltage bin. The frequency bins will change based on thermals, but it’ll try to use 5 bins higher.
My anecdotal evidence for this is that you can set your curve while hot and if you subsequently cool the card and then run under load, you’ll see the frequency go higher than what you saw when you set the curve, e.g. maybe 900mV was 1830 when you set that as the limit of your curve, but then it’ll boost to 1875 at 900mV after you let it cool and re-load the GPU without touching the curve.
Hello, everyone...just getting here and trying undervolting on my 3090 ventus. I notice that I can set my curve so that it barely moves from the max clock during timespy and superposition but will fluctuate between my boost clock (1725) and my top clock speed quite a bit during games like gta5 and modded skyrim sse at 4k ultra settings. For example, in timespy and superposition I can do 1845mhz @ .812 with barely any power throttling and only minor clock fluctuations from like 1815 to 1860. But during games, where I don't power throttle at all at the same settings, the clocks will go from 1725-1860. Any ideas what might be going on? When I run at a constant 1725 @ .775 everything seems to be much more smooth in games and the clock fluctuations only range from 1725-1740. Any help understanding what's going on would be greatly appreciated!
waderunner
[H]ard|Gawd
- Joined
- Jun 2, 2013
- Messages
- 1,130
Very helpful discussion. Undervolting really helped me improve coil whine on my EVGA 3070 XC3. It does seem as if the frequency / power curve is pretty heavy handed. Backing off the power by undervolting really minimizes the coil whine on my card to the point that it isn't an issue (can't really hear it through the case). And, I'm not seeing any impact in 3DMark scores. Win-Win. Initially thought I'd have to return the card. Thanks to all sharing the knowledge.
dr.stevil
[H]F Junkie
- Joined
- Sep 26, 2008
- Messages
- 9,186
I just decided to try undervolting my 3090 hybrid last night and I’m amazed with the results. I honestly don’t know why I didn’t try this sooner. Temps in some cases are as much as 10* lower with the same, and in some cases more, performance.
Win win!
Win win!
pippenainteasy
[H]ard|Gawd
- Joined
- May 20, 2016
- Messages
- 1,137
And with lower temps come lower fan speed. So it's a real quality of life issue. Maybe the guys with custom loops don't care that much, but it helps a ton on air.
Finally had time to play around with this, .9mv @1900mhz works great with 104%+ power target. Have not fine tuned it yet. .85mv and 1850mhz is also good so far. I am talking almost an actual 200mhz increase in clock speed over default in games which I think is rather decent.
Not sure if you are just rounding in the post or not. But try and keep the clock speeds in 15 Mhz increments as that is the increments that the GPU uses. Your card seems to be close to mine. I’ve also been successful with 1950 @ .925 and 2010 @ .950 as well; 100% rock stable in all games.
An EVGA AIO is totally worth installing. I have one on my 3080 XC3 Ultra and it is fantastic.
thanks for the detailed Guide...I recently got my hands on an EVGA 3080 FTW3 Ultra and keep hearing people recommend undervolting it...I've never undervolted a GPU so I'm wondering does this only effect people manually overclocking the card?...if I leave the FTW3 Ultra at stock settings will it most likely be fine as far as voltages and power draw?
the FTW3 Ultra is factory overclocked and still uses those Max Boosts but will leaving everything at those stock settings be good enough or does it still require an undervolt?...any risks to undervolting as far as damage to the card itself, performance etc?...or does it pose the same risks as overclocking?
the FTW3 Ultra is factory overclocked and still uses those Max Boosts but will leaving everything at those stock settings be good enough or does it still require an undervolt?...any risks to undervolting as far as damage to the card itself, performance etc?...or does it pose the same risks as overclocking?
if you leave it stock, you will absolutely hit tdp when you use it and be power limited regardless of if you undervolt or not. An undervolt is if you want more performance or less power, or both. Undervolting is essentially overclocking at a given voltage, and carries the same risks, which are few. You aren't so much turning down voltage as you are turning up clock speed at a given voltage, so that when the card says run at ~2000mhz, you are able to do so at a lower voltage, resulting in less heat and power.
Undervolting will allow for lower voltage (sure you figured this out!) for the same clock thus reducing heat and lower fan. Like Steve Jobs would have said : it's magic!
I saw a video detailing undervolting the 3080 card I currently have (FTW3 Ultra) so I might just use the same numbers- 850mV and 1815MHz
isn't the stock FTW3 Ultra 1800MHz?...so all I'm really doing is lowering the voltage while keeping everything else the same?
isn't the stock FTW3 Ultra 1800MHz?...so all I'm really doing is lowering the voltage while keeping everything else the same?
Hi
i've undervolted my Strix 3090 to 0.850 @ 1905 but i find my temps quite hot even with it
while playing some games like days gone, i have around 70c on core and 86c on memory
I game at 3840x1600, i've also put some heatsink and a fan on the backplate to cool vrm but not sure if it does something to be honest
I don't want to go on water for now or change thermal pads for warranty issues
Are thos ok temps for this gpu? I'm quite concerned by temps
If i go lower like 1850 for 0.8 will I lose a lot of performances?
Thanks
i've undervolted my Strix 3090 to 0.850 @ 1905 but i find my temps quite hot even with it
while playing some games like days gone, i have around 70c on core and 86c on memory
I game at 3840x1600, i've also put some heatsink and a fan on the backplate to cool vrm but not sure if it does something to be honest
I don't want to go on water for now or change thermal pads for warranty issues
Are thos ok temps for this gpu? I'm quite concerned by temps
If i go lower like 1850 for 0.8 will I lose a lot of performances?
Thanks
temps look fine. vram by design doesn't throttle until 108c.
I couldn't find a straightforward answer googling this.
Can sliders that aren't a multiple of 25 mV be adjusted on their own?
GPU is an air cooled TUF 3070 TI OC Edition in a Torrent Compact.
This is stock at 33C, although sometimes it boosts to 2025 and 2040:
This is my current UV (+1200 memory) using OP's method, bar forcing a lower top clocks. Power limit is untouched at 100%.
This is stable, though I haven't done extensive stress testing. The games I play run fine. No crashing no artifacting. Time spy and Heaven are also ok. Haven't ran BMI.
1980@925 results in Apex Legends crashing after 1-2 minutes of in-game time.
I want to try 1980@943 and if that's stable 937 and 931. However after setting the slider (or any of the sliders that aren't multiples of 25) and clicking apply the graph just resets to what I started with.
Am I doing something wrong or is that generally not possible?
Is there a better way of further undervolting from where I've gotten?
Can sliders that aren't a multiple of 25 mV be adjusted on their own?
GPU is an air cooled TUF 3070 TI OC Edition in a Torrent Compact.
This is stock at 33C, although sometimes it boosts to 2025 and 2040:
This is my current UV (+1200 memory) using OP's method, bar forcing a lower top clocks. Power limit is untouched at 100%.
This is stable, though I haven't done extensive stress testing. The games I play run fine. No crashing no artifacting. Time spy and Heaven are also ok. Haven't ran BMI.
1980@925 results in Apex Legends crashing after 1-2 minutes of in-game time.
I want to try 1980@943 and if that's stable 937 and 931. However after setting the slider (or any of the sliders that aren't multiples of 25) and clicking apply the graph just resets to what I started with.
Am I doing something wrong or is that generally not possible?
Is there a better way of further undervolting from where I've gotten?
Last edited:
LittleBuddy
Limp Gawd
- Joined
- Jan 3, 2023
- Messages
- 174
Are these crashes nvlddmkm in event viewer? If so it's unstable. If not it could be something else.
They are indeed nvlddmkm events. But the crashes are not what I'm trying to troubleshoot. It's the Afterburner curve sliders not staying where I want them to.
LittleBuddy
Limp Gawd
- Joined
- Jan 3, 2023
- Messages
- 174
On the V/F curve you can adjust the voltage and frequency at any point on the curve up to the maximum voltage supported on the card. I'm not entirely sure why the OP suggests 25mV increments, maybe for simplicity. But if you want to do a simple undervolt, just use the stable overclock you have on your card and cap the voltage beyond that point. So if your card is stable through the entire curve at +150 and you want to cap your voltage at 950mV just flatten the curve beyond 950mV. If you want to use the power limit slider, use it sparingly, because it has a large impact on GPU performance. An Example would be here https://hardforum.com/threads/anyone-undervolting-the-4090.2024408/#post-1045550183 for 40 series but the same applies for 30 series cards.
Here's an example of a +150 OC with a voltage limit of 950mV:
On 30 series more so than 40 series the curve is temperature dependent as well so you will want to apply your settings at the same temperature every time, personally I did it at 30C after every boot.