note: Dragging down the power slider will be absolutely fine. Performance loss there will depend on the game. I tried a 75% power limit and saw clocks drop to 2550MHz (-200MHz). Also, the video is not clickbait. I legitimately don't recommend manually adjusting the vcore for this GPU. I can't explain why there's a performance loss when the core clock is same / higher than stock. Those saying "only 3% performance loss @ -70W" are missing the point. I've reached out to Nvidia for more insight on what's happening in the background here and will update you with their response.
Yeah, power limit is the best way to go if you have thermal issues or weaker psu. Also don't forget that undervolting is overclocking, as usual it will vary per unit.
In der8auer's 11 Oct video at about the 14-1/2 minute point, he starts talking about the power draw. He didn't even do an actual undervolt (that didn't work well). All he did was limit the Power Target. 33% power reduction for 5% (i.e., indistinguishable) performance loss.
Effective clock is dropping because you are using the incorrect way of undervolt. DON'T MOVE THE GRAPH DOWN. Dots down = overvolt/underclock. Dots up = undervolt/overclock. By moving the graph down you are decreasing performance when voltage fluctuates. (IE why effective clock is dropping.) How to undervolt correctly: 1. Select desired max voltage dot. 2. Move the graph UP with shift+mouse by desired frequency. 3. Select the right side of graph (including the dot) with shift+mouse. 4. Press shift+enter 2 times. That way the effective clock won't drop. (The similar behaviour on previous generations, just the difference with effective clock isn't that big as with ada).
I see Soo many RUclipsrs moving the graph down and then picking the target voltage x clock speed and just raising that single point and capping it. I don't get why they initially shift the graph down. That always made no sense to me.
This is why you always need to validate overclocks with benchmarks. Clock speeds don't tell you whether your overclock is either stable or actually increasing the effective framerates.
Try to raise *whole* freq./volt. curve up, all points at once. The plot shift will be smaller but the real clocks will match the effective ones under all power scenarios. Do not deceive yourself with rising just the last point of curve to the desired clock speed. Remember: undervolting is like an overclocking under lower voltages.
But by raising that last points, all points on the left side will be automatically lifted up, when he clicked safe, so i dont get what you are saying and the results speak for itself?
this seems like exactly the same thing that happens with ryzen cpus on certain scenarios, they call it CLOCK STRETCHING. This "feature" is meant to increase stability and I think there`s no way to turn it off.
Someone mentioned that the way he undervolted it was wrong. Moving the entire graph down is the reason that the other clock speeds lowered along with it which kinda makes sense as to why power limiting didn't affect those other clock speeds.
Seems way worth it!!! Save 100 watts, like 20-25% for only few points loss on already ludicrous performance!! Absolutely worth it, but interesting how this doesn't work like normal. Odd!
It sounds to me like they may have clock-stretching implemented, meaning it will dynamically under-clock when not enough voltage is available. However, all power saving are benefitial, so find the largest under-volt that does not result in this behaviour and then move on to other tuning options.
Yeah, seems to be the trend. My 12600k also clock-stretches below -60-70mV. It's a bit annoying because you have to compare benchmark scores the whole time. In the past you just had to see if your system crashes or not, lol
This video honestly needs to be deleted. It's bad advice based on misinformation. The core issue here is that you have incorrectly applied the undervolt.... The entire curve needs to be lifted up and then flattened beyond the target, which preserves the shape of the curve. Not like the way it's done here. Your aim should be to offset the existing curve in its entirety, so that all voltage/clock points are shifted relative to each other. The way you've done it here has completely reshaped the curve, and made a horribly steep and linear increase. You've also set voltage levels that are drastically different than stock. Furthermore, core clock and effective clock are measuring the same thing, just with different sampling logic. The core clock is the highest recorded value within a given time period, whereas the effective clock is the average value recorded over that same period. Effective will always be lower than core, but a big difference between the two indicates core clock instability, or variance from bumping into thermal or power limits. In this case, you're getting instability at the top end because your curve is done wrong. Your actual core clock is fluctuating wildly because the voltage points are irregular. Lastly, the Video clock is also meaningless for games. It comes from the video engine that is used for hardware video decoding. Just ignore it, it's unimportant.
@@haven216The correct way to apply an undervolt is to first raise the entire curve by shift clicking, and then dragging the entire curve up by a desired amount. I.e lifting the entire curve, as opposed to the way it’s done here where he lifts only a single point. If you want to see the proper way to do this, then have a look at the video called ‘Undervolt your RTX 4090 in THE RIGHT WAY’ by a guy called ‘ImWateringPSUs’
5:22 in the Video, so you lose 1-2% fps. At the same time you draw 17% less power. So why again shouldnt i undervolt? This is an efficiency increas of 16%! I think your own data shows the opposite of what you are concluding. The ratio of power to fps is the important thing and not the fact, that you lose 5% performance when you undervolt, because, again ,you draw much less power.
I'd be the kind of guy to limit framerates in many games on this GPU just to keep powerconsumption down while still reaching insanely fast frame timing
From what I've seen you usually get better efficiency by lowering the power limit over using an fps limit. Maybe you get worse input latency or something but nvidia has it's low latency mode that helps in cases where the gpu load is close to 100% so it isn't a big deal.
Der 8auer tested it and the 4090 is completely voltage constraint, even when increasing the power target. So lowering the power targer even 30-40% results in more than hundred watts less with almost no performance decrease in gaming.
It's not really undervolting if you lose performance because of it. The whole point of undervolting is to lower power consumption and keep clockspeed the same
@@stronk9969 it’s because nvidia is pushing the card so far outside efficiency window, which was usually left for OC’ers, that would take 100 more watts for 5% more performance.
Nvidia probably already used that trick (I mean to undervolt) to some extend to minimize the power consumption so you are left with considerably less margin than before. It's the same thing as with CPU in the early 2010s era you could overclock easily, nowadays there isn't that much room left.
@@Killer344-ZxS that isn’t binning bruh. That’s just factory optimization. Binning is when they make and sell multiple tiers of the same product based on quality. RAM is the best example of binning. It’d be like if they had 3000mhz 4090s available starting at $2000
The point he’s trying to make is that they still have to set standard voltage values that overshoot to accommodate for some variation in the quality of the GPU’s that come out so by that understanding there’s no reason why it shouldn’t still be valid
@@ChiquitaSpeaks here I thought that was a foregone conclusion, that overclocking/changing voltages specific to the silicon be performed post-purchase even if one does not intend to overclock - since the end user can eek out the limits of the silicon they got and the factory doesn't have that level of individual attention. It makes sense to me, but I guess this WILL be the first card I actually saw fit to undervolt - not that the 3080 I was using wasn't a power hog too - not quite on the same level as this monster though.
I undervolted my 4070ti. Only tested in CS2, but my temps went down from 50 to 35 under load, and Power consumption went from 120W to 60-70W. Didn't notice any FPS difference. Definitely very nice, but I wonder how it will behave in more demanding games.
Lol looks fine. In fact I’d say performing better than power limiting performance vs power draw from just pure stats. There was a very small frequency drop w performance in the power limiting video with Der8auer too if there was one at all, plus since you didn’t show the stock 4090 video/effective clocks I couldn’t really make out the comparison you presented well in this video. If anything i’m thinking the voltage changes limiting factor would be the new cache system in these GPU’s. Btw have you tried doing side by side undervolt vs power limit & undervolt…you should give it a try?
Have you tried overclocking with slider + lowering power limit? That's equivalent to undervolting (as you're raising all points of voltage/frequency curve and then limit max. frequency back to stock with PL => get stock freq at lower voltage).
This is what should be happening. You shouldn't expect to run less power and maintain the same performance or gain it. The beauty in undervolting is to reduce power consumption and heat load in your room while losing maybe 5% performance.
This is extremely similar behavior to Ryzen clock stretching. The clocks might appear high, but without enough voltage the performance suffers. This just shows that TSMC 5nm node scales really well with voltage, and probably why NVIDIA claimed that their H100 accelerators can scale to up to 800W.
Interesting yours capped at 0.9v, mine can't go below 0.86v (MSI 4090 SUPRIM X model), but I don't notice that lower fps from the undervolting (e.g running 2760-2775 @ 1.05v default vs 0.95v gives similar performance) - i wonder if there is some kind of bin difference on the GPUs
@@optimumtech it reminds me of ryzen cpus, sometimes with UV it gives lower scores even on same frequencies. I think its anti-bsod error correcting thing which wont allow your pc crash. I get lower fps from UV on my msi trio 3080, even at same frequency. UV is still best way to go
@@optimumtech I checked with 3DMark TSE, The effective clock is similar on default vs undervolted here, default is 2775-2760 @ 1.05v, effective clock 2720-2730Mhz. Undervolted at 2775 950mV, effective clock still show 2730, not dropped to 26xx Mhz range like yours.
@@m7thena It's too early to tell though, i also only have one sample on hand for my review and cannot conclude stuff based on this one sample - it might be there's a sample-to-sample variation from either FE or AIB units. Actually we may need more people testing these stuff so we know better :)
Thanks for the video! This informative benchmark/ hardware/ software type video in under 10 minutes is 1 of the things I LOVE most about you and your channel. -Alan
Because you way of undervolting and editing the voltage curve is simply wrong and you did not run benchmark like timespy extreme to make sure it is actually stable under heavy load. Some games still can run even the GPU is not 100% stable.
Derbauer showed that the RTX 4090 is voltage limited. If you use gpu-z you will see this. Out of the box the gpu is voltage limited so reducing voltage is going to reduce performance. Adjusting the power limit is the only thing you can really reduce.
Have you tried strictly power limiting the card rather than under volting it? The secondary clocks are clearly tied to voltage in this case which is not the same as the 3090.
What's happening here is that the GPU turns parts of itself off more often and submits less threads to execution. Power Gating and Thread Queue stalling. Thereby keeping less GPU resources occupied while making do with the reduced power budget. This is built in 'by design' behaviour.
what? setting a specifc voltage vs dropping the entire curve are different. If you hold shift and drag click, select all voltages past the voltage you want to set it to. then drop the highest point below that voltage (lazy way is to just spike it directly down). You keep the memory clock WHILE UNDERVOLTED.
I would change the title, still seems worth undervolting considering the high power usage, and I'd rather lower overall power usage if I can even if it costs me 3-5% performance, I won't miss it and my power bill will be much happier.
I don't think this right. ImWateringPSUs did a video explaining how to undervolt the 4090 - /watch?v=WjYH6oVb2Uw - and the results are amazing. You can run at the boost clock while undervolted @975 while also applying an OC to the memory to achieve the exact same performance as stock but while using using significantly less power. Dropping power has shown to cause issues and by undervolting you are avoiding these issues by fixing the clock at a specific speed/voltage, so games should and do run smoother.
These results aren't actually all that surprising. Several channels have shown the 4090 is voltage limited, not power limited, so reducing the voltage should absolutely reduce performance on that card.
You lose performance lowering voltage at the same frequency when its unstable. But manufacturers don't rly want to be bothered with unstable cards so they just clock stretch it. Looks like its having the same frequency but it actually loses performance (less core stress) I don't know by any means how that actually works but it definitely has to do with stability matter and the fact that no one wants to see instability at any circumstance from either side. Customer complaint and driver tweak to fix it on every single game title.
Hello. If you applied the curve on the graph as shown in the video, then you didn't do undervolting but overvolting, and that's why there is performance degradation of the graphics card.
Since you never show the 4090 UV curve, did you tune ALL the voltage points below the target volt/frequency point instead of having the cliff? On my 2080ti the difference of tuning those at 1860Mhz target was effective clock going from 1780 to 1840. I'm kinda surprised you're even getting that high of an effective clock on the 3090 with that curve.
I don't get the point being made here, performance is slightly lowered from stock settings, 4-10 fps , but decreases power consumption by like 80 watts and reduces heat. I think ill keep under-volting.
It's the best thing for many years, always preferred this to undervolt. The best thing to do is core oc + memory oc + temp limit (usually 65 degrees or lower, so as to have hot spots and memories at 75 degrees maximum in gaming and under stress, not mining. In short, adjust the temperature so as to have memories and hotspots at no more than 75 degrees). So the PL is based on temperature but you have to remember to choose to prefer the temp limit to the PL in MSI afterburner.
It's the same with modern CPUs. Rather than hitting PLs, set the temperature limit on both AMD and Intel (Asus board only for Intel, it's an exclusive AI Cooling feature) and add negative adaptive voltage for Intel or curve optimizer for AMD.
Despite the clickbaity title of the video, you can still undervolt AND overclock your 4090 at the same time and you get better results. PS: Afterburner will probably get an update improving the overclocking capabilities and voltage regulation of the new GPUs as with every early release.
Quite useful. I love your unique content and they exactly matches my interests. I was wondering what happens if we overclock and undervolt? Does the effective clock go up or not? I was also wondering what is the power draw on 0.905v lowest possible. How much efficient this thing can go against the 4080s?
So you get a little less performance at the same clock speed when undervolting - but also less powerdraw on the other side you get less performance (less clockspeed) at the same Power level when reducing the Power Target. It's all about your priority: are you going to stay below a certain power level or are you going to stay above a certain performance level?
You mentioned that you can't go below 0.9 V and never saw that behaviour before. Actually it's the same with the 3090 Ti, I can't go below 0.9 V on my MSI RTX 3090 Ti Suprim X.
Nah! I love undervolting my 4090 to the absolute MAX! and it's amazing! UV at 875mV and 2500Mhz the performance is massive and I even use FPS limit AND DLSS!!! So it's the maximum possible to save very much power but still have more than enough performance in gaming on 4k@120Hz! In the most games still I get my over 100 FPS and it's absolutely amazing ❤
same. UVed at 900mV and 2550Mhz + 141fps limiter (144fps - 3fps Headroom for possible framespikes). Power Consumption went down from 439W peak (Borderlands 3 Benchmark, 1440p, Max settings + Reshade) to 272W peak, with temps barely pass 45 degree Celcius, while maintaining constant 141fps.
Still consider an undervolt if your trying to reuse say your old 750 watt power supply.Couple the undervolt with say an older 65 watt CPU and you should be able to do that ! Especially if you also consider those reported transient power spikes in the GPU's !
der8auer found that just reducing the power target was the best option instead of undervolting
Just wanted to write the same
I've found the same with my 3080ti
Correct. Try something like 70-80% power target + 100-150Mhz on the core and see what kind of performance you'll get
I do the same with my 2070 Super. It not only lowers temps and lower fan speeds, but gets rid of coil whine too with only ~3fps loss on average.
Probably for the best tbh - but you're still looking at a core-clock reduction and performance loss.
note: Dragging down the power slider will be absolutely fine. Performance loss there will depend on the game. I tried a 75% power limit and saw clocks drop to 2550MHz (-200MHz).
Also, the video is not clickbait. I legitimately don't recommend manually adjusting the vcore for this GPU. I can't explain why there's a performance loss when the core clock is same / higher than stock. Those saying "only 3% performance loss @ -70W" are missing the point.
I've reached out to Nvidia for more insight on what's happening in the background here and will update you with their response.
Yeah, power limit is the best way to go if you have thermal issues or weaker psu.
Also don't forget that undervolting is overclocking, as usual it will vary per unit.
Always informative 👍 👌
It's a power hungry GPU. So it makes sense to not play with the power and make it unstable
@@riba2233 just depends on the game, just pick the one that makes more sense for you.
@@riba2233 you can undervolt and overclock at the same time so they are not same actually
Hmm, after all 70W power reduction (about 18%) for 2% perf. decrease is still worth it IMO.
In der8auer's 11 Oct video at about the 14-1/2 minute point, he starts talking about the power draw. He didn't even do an actual undervolt (that didn't work well). All he did was limit the Power Target. 33% power reduction for 5% (i.e., indistinguishable) performance loss.
The performance loss shown in this video is negligible as well. So why not undervolt, not seeing the downsides to justify the header.
its been this way for generations since maxwell at least
@@HeloisGevit Header is stupid, but you are much better off just decreasing power target.
@@Navhkrin why is it "much better"?
@@GGYlenol You can run into stability issues with undervolt, no such issues with power target.
Effective clock is dropping because you are using the incorrect way of undervolt.
DON'T MOVE THE GRAPH DOWN.
Dots down = overvolt/underclock.
Dots up = undervolt/overclock.
By moving the graph down you are decreasing performance when voltage fluctuates. (IE why effective clock is dropping.)
How to undervolt correctly:
1. Select desired max voltage dot.
2. Move the graph UP with shift+mouse by desired frequency.
3. Select the right side of graph (including the dot) with shift+mouse.
4. Press shift+enter 2 times.
That way the effective clock won't drop. (The similar behaviour on previous generations, just the difference with effective clock isn't that big as with ada).
I see Soo many RUclipsrs moving the graph down and then picking the target voltage x clock speed and just raising that single point and capping it. I don't get why they initially shift the graph down. That always made no sense to me.
The video clock also drops with this method you are mentioning. From 2100 to 1965 for my Manli RTX 4090 at 2730@950
thanks for pointing out this guy has no idea what he’s doing. you are correct
he technically did it correctly at the start which is strange
This is why you always need to validate overclocks with benchmarks. Clock speeds don't tell you whether your overclock is either stable or actually increasing the effective framerates.
Try to raise *whole* freq./volt. curve up, all points at once. The plot shift will be smaller but the real clocks will match the effective ones under all power scenarios. Do not deceive yourself with rising just the last point of curve to the desired clock speed.
Remember: undervolting is like an overclocking under lower voltages.
Exactly. I don't know why so few people noticed the borked curve.
aka do normal overclocking with lower power limit
that's exactly what I always do
But by raising that last points, all points on the left side will be automatically lifted up, when he clicked safe, so i dont get what you are saying and the results speak for itself?
this seems like exactly the same thing that happens with ryzen cpus on certain scenarios, they call it CLOCK STRETCHING. This "feature" is meant to increase stability and I think there`s no way to turn it off.
Happens to 12th gen intels as well, or at least with my 12600k. Below -70mV I start getting lower R23 scores...
Someone mentioned that the way he undervolted it was wrong. Moving the entire graph down is the reason that the other clock speeds lowered along with it which kinda makes sense as to why power limiting didn't affect those other clock speeds.
i would still do it. margin of error differences for 100w less seems worth it
exactly. i don't even understand wtf is this video about. you gain 100w and I'm sure 5/6 degrees on a 5/10 fps loss.
@@davidepannone6021 This video is exclusively about clicks, nothing else.
Seems way worth it!!! Save 100 watts, like 20-25% for only few points loss on already ludicrous performance!! Absolutely worth it, but interesting how this doesn't work like normal. Odd!
Bot💀
@@Wxltzy5陰火牛 i keep reporting this but for some reason it stays there
It sounds to me like they may have clock-stretching implemented, meaning it will dynamically under-clock when not enough voltage is available. However, all power saving are benefitial, so find the largest under-volt that does not result in this behaviour and then move on to other tuning options.
Yeah, seems to be the trend. My 12600k also clock-stretches below -60-70mV. It's a bit annoying because you have to compare benchmark scores the whole time.
In the past you just had to see if your system crashes or not, lol
beneficial*
This video honestly needs to be deleted. It's bad advice based on misinformation.
The core issue here is that you have incorrectly applied the undervolt.... The entire curve needs to be lifted up and then flattened beyond the target, which preserves the shape of the curve. Not like the way it's done here. Your aim should be to offset the existing curve in its entirety, so that all voltage/clock points are shifted relative to each other. The way you've done it here has completely reshaped the curve, and made a horribly steep and linear increase. You've also set voltage levels that are drastically different than stock.
Furthermore, core clock and effective clock are measuring the same thing, just with different sampling logic. The core clock is the highest recorded value within a given time period, whereas the effective clock is the average value recorded over that same period. Effective will always be lower than core, but a big difference between the two indicates core clock instability, or variance from bumping into thermal or power limits. In this case, you're getting instability at the top end because your curve is done wrong. Your actual core clock is fluctuating wildly because the voltage points are irregular.
Lastly, the Video clock is also meaningless for games. It comes from the video engine that is used for hardware video decoding. Just ignore it, it's unimportant.
What do you mean by lifting up the entire curve?
@@haven216The correct way to apply an undervolt is to first raise the entire curve by shift clicking, and then dragging the entire curve up by a desired amount. I.e lifting the entire curve, as opposed to the way it’s done here where he lifts only a single point.
If you want to see the proper way to do this, then have a look at the video called ‘Undervolt your RTX 4090 in THE RIGHT WAY’ by a guy called ‘ImWateringPSUs’
5:22 in the Video, so you lose 1-2% fps. At the same time you draw 17% less power. So why again shouldnt i undervolt? This is an efficiency increas of 16%!
I think your own data shows the opposite of what you are concluding. The ratio of power to fps is the important thing and not the fact, that you lose 5% performance when you undervolt, because, again ,you draw much less power.
He just likes Clickbait titles.
Thanks for covering this. Your undervolt guide for the 3090 was really helpful and it's good to know what's the deal this generation.
I'd be the kind of guy to limit framerates in many games on this GPU just to keep powerconsumption down while still reaching insanely fast frame timing
Fr tbh, you only need just enough frames most of the time but the highest quality of image
From what I've seen you usually get better efficiency by lowering the power limit over using an fps limit. Maybe you get worse input latency or something but nvidia has it's low latency mode that helps in cases where the gpu load is close to 100% so it isn't a big deal.
@@FakeMichau doesn't matter, having perfect frame rate consistency is more beneficial than 50 frames more.
Vsync adds input lag your loss lmao
@@OMGERRORWTF Noone said anything about vsync. I limit via nvidia controls
Der 8auer tested it and the 4090 is completely voltage constraint, even when increasing the power target. So lowering the power targer even 30-40% results in more than hundred watts less with almost no performance decrease in gaming.
Just lower power target, it has same performance at 90% PT and at 70% it only loses 5% fps
Nice
It's not really undervolting if you lose performance because of it. The whole point of undervolting is to lower power consumption and keep clockspeed the same
@@stronk9969 yeah and that was the point of the video
@@Serexityy_ Exactly
@@stronk9969 it’s because nvidia is pushing the card so far outside efficiency window, which was usually left for OC’ers, that would take 100 more watts for 5% more performance.
Nvidia probably already used that trick (I mean to undervolt) to some extend to minimize the power consumption so you are left with considerably less margin than before.
It's the same thing as with CPU in the early 2010s era you could overclock easily, nowadays there isn't that much room left.
That is called binning lol.
@@Killer344-ZxS that isn’t binning bruh. That’s just factory optimization. Binning is when they make and sell multiple tiers of the same product based on quality. RAM is the best example of binning. It’d be like if they had 3000mhz 4090s available starting at $2000
The point he’s trying to make is that they still have to set standard voltage values that overshoot to accommodate for some variation in the quality of the GPU’s that come out so by that understanding there’s no reason why it shouldn’t still be valid
@@Killer344-ZxSno? gpus are not binned, they use different chips
@@ChiquitaSpeaks here I thought that was a foregone conclusion, that overclocking/changing voltages specific to the silicon be performed post-purchase even if one does not intend to overclock - since the end user can eek out the limits of the silicon they got and the factory doesn't have that level of individual attention. It makes sense to me, but I guess this WILL be the first card I actually saw fit to undervolt - not that the 3080 I was using wasn't a power hog too - not quite on the same level as this monster though.
You should undervolt by overclocking the core to desired frequency at certain voltage and then flatten the curve by holding ctrl and swiping down
I undervolted my 4070ti. Only tested in CS2, but my temps went down from 50 to 35 under load, and Power consumption went from 120W to 60-70W. Didn't notice any FPS difference. Definitely very nice, but I wonder how it will behave in more demanding games.
Lol looks fine. In fact I’d say performing better than power limiting performance vs power draw from just pure stats. There was a very small frequency drop w performance in the power limiting video with Der8auer too if there was one at all, plus since you didn’t show the stock 4090 video/effective clocks I couldn’t really make out the comparison you presented well in this video. If anything i’m thinking the voltage changes limiting factor would be the new cache system in these GPU’s. Btw have you tried doing side by side undervolt vs power limit & undervolt…you should give it a try?
Thank you for your constant hard work and investigations.
He’s not going to notice you
@@VaydaladaVodalada Who fucking hurt you dude
Thank you!
@@optimumtech lmao !
@@VaydaladaVodalada Rude
Have you tried overclocking with slider + lowering power limit? That's equivalent to undervolting (as you're raising all points of voltage/frequency curve and then limit max. frequency back to stock with PL => get stock freq at lower voltage).
I need to try this! Looks very promising
This is what should be happening. You shouldn't expect to run less power and maintain the same performance or gain it. The beauty in undervolting is to reduce power consumption and heat load in your room while losing maybe 5% performance.
my 3080 performed better when undervolted,same my cpu. the lows are often way better.
Try to test it urself. Tons of examples out there with UV + OC on 4090. U can get better perf with less power draw
I can clock mine @ 2775 MHz with 1500 memory OC @ 950mv. 3% better than stock with 400W power draw max in synthetic benchmark
@@kellen4854 The key word in my comment is "expect". Not trying to say you can't chase performance and efficiency with your own card.
Don't? Losing 3 fps to reduce by 70 watt is clearly worth it.
People should undervolt anyway, just to consume less power on the grid.
This is extremely similar behavior to Ryzen clock stretching. The clocks might appear high, but without enough voltage the performance suffers. This just shows that TSMC 5nm node scales really well with voltage, and probably why NVIDIA claimed that their H100 accelerators can scale to up to 800W.
Interesting yours capped at 0.9v, mine can't go below 0.86v (MSI 4090 SUPRIM X model), but I don't notice that lower fps from the undervolting (e.g running 2760-2775 @ 1.05v default vs 0.95v gives similar performance) - i wonder if there is some kind of bin difference on the GPUs
How about your video clock and effective clock - do they auto-regulate like shown here?
@@optimumtech it reminds me of ryzen cpus, sometimes with UV it gives lower scores even on same frequencies. I think its anti-bsod error correcting thing which wont allow your pc crash. I get lower fps from UV on my msi trio 3080, even at same frequency. UV is still best way to go
@@optimumtech I checked with 3DMark TSE, The effective clock is similar on default vs undervolted here, default is 2775-2760 @ 1.05v, effective clock 2720-2730Mhz. Undervolted at 2775 950mV, effective clock still show 2730, not dropped to 26xx Mhz range like yours.
@@Luckyn00bOC This puts @Optimum Tech 's review in doubt. I don't see the point of this video and the blunt title misleading viewers.
@@m7thena It's too early to tell though, i also only have one sample on hand for my review and cannot conclude stuff based on this one sample - it might be there's a sample-to-sample variation from either FE or AIB units.
Actually we may need more people testing these stuff so we know better :)
I think undervolting the 4090 deserves another video.
I would totally take 3 to 5% performance penalty for 100W less power draw. You don't know how crazy electricity prices are in Europe!
Just lowering the power limit in afterburner seems to work
Der8auer had some different findings.
Thanks for the video! This informative benchmark/ hardware/ software type video in under 10 minutes is 1 of the things I LOVE most about you and your channel. -Alan
Maybe it gets better with driver updates or something? Because I LOVE undervolting :( I mean I would risk that 10 fps or something but..... Yeah
Same. Like I'd rather have 3-5% performance hit than this card pulling 400w off the wall all the time. That's just insane.
That would be great!
@@ericzedd You can still get that by power limiting to 70% as per der bauer.
@@romxxii Thanks for that!
The "full stock" tuning is so far past reasonable efficiency that we might as well regard it as already pre-overclocked
It's weird not to suggest reducing power consumption when the FPS loss is only 1 to 5 FPS at an average of over 100FPS.
Because you way of undervolting and editing the voltage curve is simply wrong and you did not run benchmark like timespy extreme to make sure it is actually stable under heavy load. Some games still can run even the GPU is not 100% stable.
After one year I can confirm this, this guy doesn't really understands what he's doing.
massive decrease in performance yeah, 456 fps becomes 440. M A S S I V E
Derbauer showed that the RTX 4090 is voltage limited. If you use gpu-z you will see this. Out of the box the gpu is voltage limited so reducing voltage is going to reduce performance. Adjusting the power limit is the only thing you can really reduce.
Awesome video man. Undervolting has been one of my favorite topics to discuss with the recent generation of GPUs coming out.
iirc i saw somewhere yesterday that u can drop to 190W and still having 90% of gaming perfomance
Have you tried strictly power limiting the card rather than under volting it? The secondary clocks are clearly tied to voltage in this case which is not the same as the 3090.
What's happening here is that the GPU turns parts of itself off more often and submits less threads to execution. Power Gating and Thread Queue stalling.
Thereby keeping less GPU resources occupied while making do with the reduced power budget. This is built in 'by design' behaviour.
Another bit that's happening is clock stretching. Look it up.
Really glad tuning an UV is dead for this gen. Rather just be able to move the power slider without much perf loss like how it seems now.
This video is out of date and incorrect. 4090 does great with an undervolt. 2850mhz @ 1v gives higher than stock perfromance with less power use.
This GPU is mostly voltage limited, not power limited. Undervolting works great for power limits but not with voltage limits.
cries in SFF
4080 is the card for most people.
@@atrium12 4080 is trash
@@atrium12 Check Daniel Owen's latest video on the 4080
@@Dankuzmeemusmaximus 4080 16GB** not the other one.
@@atrium12 Both are trash
0:33 what an epic shot!
how'd you open the box though? string attached? or some rotoscoping :D
Just standing really far back and lifting up the other side lol
BTW, it's the same for Ryzen 5000. Clock speed is the same, but performance will drop.
this is not done correctly
Just reduce the power limit! Like in the 3080, 3090... And +XXX core +XXX memory
Wouldn't the 1% fps lows improve with increased stability if undervolting
in superposition my lows stock is 70 fps and 100 fps undervoltet
compared to ?@@ReDHaWko
UV 1.0v
99% perf
-3 degree
- 46w wasted
Lesser noise
what? setting a specifc voltage vs dropping the entire curve are different. If you hold shift and drag click, select all voltages past the voltage you want to set it to. then drop the highest point below that voltage (lazy way is to just spike it directly down). You keep the memory clock WHILE UNDERVOLTED.
I would change the title, still seems worth undervolting considering the high power usage, and I'd rather lower overall power usage if I can even if it costs me 3-5% performance, I won't miss it and my power bill will be much happier.
Lower power draw to 60-70% instead it's more efficient then undervolting
I don't think this right. ImWateringPSUs did a video explaining how to undervolt the 4090 - /watch?v=WjYH6oVb2Uw - and the results are amazing. You can run at the boost clock while undervolted @975 while also applying an OC to the memory to achieve the exact same performance as stock but while using using significantly less power. Dropping power has shown to cause issues and by undervolting you are avoiding these issues by fixing the clock at a specific speed/voltage, so games should and do run smoother.
These results aren't actually all that surprising. Several channels have shown the 4090 is voltage limited, not power limited, so reducing the voltage should absolutely reduce performance on that card.
You lose performance lowering voltage at the same frequency when its unstable. But manufacturers don't rly want to be bothered with unstable cards so they just clock stretch it. Looks like its having the same frequency but it actually loses performance (less core stress)
I don't know by any means how that actually works but it definitely has to do with stability matter and the fact that no one wants to see instability at any circumstance from either side. Customer complaint and driver tweak to fix it on every single game title.
5% performance loss, but 100W less power consumption? Damn, count me in!
Hello. If you applied the curve on the graph as shown in the video, then you didn't do undervolting but overvolting, and that's why there is performance degradation of the graphics card.
Everyone is freaking out about 450w and it cracks me up like the 3090 FTW3U and others have been doing that for years now.
You don't always need full power, so undervolt is useful.
Since you never show the 4090 UV curve, did you tune ALL the voltage points below the target volt/frequency point instead of having the cliff? On my 2080ti the difference of tuning those at 1860Mhz target was effective clock going from 1780 to 1840. I'm kinda surprised you're even getting that high of an effective clock on the 3090 with that curve.
Can't wait to see what itx builds you can manage with the 4090.
Most of the bechmsrkers I follow have noticed that gpu-z tells them that perfcap says voltage is what's holding the 4090 back so this is no surprise
2% less Perf by 20% less power consumption? Do undervolt!
Seems it needs all the voltage, but not the current. Voltage limit hits before power limit
That’s clock stretching in full effect.
Yet there’s been reports of the 12VHPWR connector melting for the four 8-PIN adapter no matter if the wires are bent or not.
I don't get the point being made here, performance is slightly lowered from stock settings, 4-10 fps , but decreases power consumption by like 80 watts and reduces heat. I think ill keep under-volting.
It's the best thing for many years, always preferred this to undervolt. The best thing to do is core oc + memory oc + temp limit (usually 65 degrees or lower, so as to have hot spots and memories at 75 degrees maximum in gaming and under stress, not mining. In short, adjust the temperature so as to have memories and hotspots at no more than 75 degrees).
So the PL is based on temperature but you have to remember to choose to prefer the temp limit to the PL in MSI afterburner.
It's the same with modern CPUs. Rather than hitting PLs, set the temperature limit on both AMD and Intel (Asus board only for Intel, it's an exclusive AI Cooling feature) and add negative adaptive voltage for Intel or curve optimizer for AMD.
do you think this will be FE specific. or will aftermarket cards react the same way?
Most likely the AIB cards will be the same.
core architecture is the same, chance are there would be no differences, unless it was related to power delivery, which doesn't seem the case anyway
@@mattgavioli6762 makes perfect sense but still sucks :/
I hope you will do some deshrouding on the other 4090 models. I wanna see some Noctua fans on them.
Despite the clickbaity title of the video, you can still undervolt AND overclock your 4090 at the same time and you get better results.
PS: Afterburner will probably get an update improving the overclocking capabilities and voltage regulation of the new GPUs as with every early release.
Afterburner won’t get an update “improving” overclocking.
@@unwinder why? they regularly add new voltage regulators support in new versions
@@JanMachovec they = me. Do not expect new versions, there is nothing to “improve”.
Half of the comments be like:DeRbAuEr AlrEaDy Found that reducing power limit most optimal thing to do....
so losing 5-8 fps isn't shit... but that power saving is well worth it!!!
I like how when i type in search "4090 fe oc", this is the first result and the second result is titled "DEFINITELY Undervolt the RTX 4090"
I feel like this means they made this card properly
Interested to see how they stack up undervolted compared to 30 series with equivalent power draw
They stacked up. It's better.
@@sakracliche Now if only Nvidia wasn't greedy
ready to lose 2% perf for 70w power reduction. In case it is not stuttering.
Quite useful. I love your unique content and they exactly matches my interests. I was wondering what happens if we overclock and undervolt? Does the effective clock go up or not? I was also wondering what is the power draw on 0.905v lowest possible. How much efficient this thing can go against the 4080s?
Interesting? Sure
Helpful? Definitnely
You're exploring areas that are rarely discussed in this detail. Thank you for sharing your experience
0:06 how did you drag all the points at once?
Good to know. Planning on cramming this into my NR200 and limiting the power to 320W.
what psu can you recommend with it for the nr 200?
@@Sonyoucef nr200 can use atx psu, so you can use whatever, atx or sfx
So you get a little less performance at the same clock speed when undervolting - but also less powerdraw
on the other side you get less performance (less clockspeed) at the same Power level when reducing the Power Target.
It's all about your priority:
are you going to stay below a certain power level or are you going to stay above a certain performance level?
just like with ampere. the real clock to check is the effective clock.
Maybe it had to do with this gen being voltage limited? at least this is the case for OC.
What I think its happening is that they already undervolted it the best they could from the factory.
for 70W... I think it is worth it.
I need ambient tenperature to make sure it lasts decades.
so just reduce the powertarget and let the GPU figure stuff out to still deliver max fps
I love gpu release season i just listen to videos all day
You mentioned that you can't go below 0.9 V and never saw that behaviour before. Actually it's the same with the 3090 Ti, I can't go below 0.9 V on my MSI RTX 3090 Ti Suprim X.
This phenomenon is called *Clock Stretching*
-10% clockspeed, but around -33% power consumption. Its insane
I thought that the 4090 had locked voltage 🤔.
The way to go from I’ve seen is by the power instead.
Nah! I love undervolting my 4090 to the absolute MAX! and it's amazing! UV at 875mV and 2500Mhz the performance is massive and I even use FPS limit AND DLSS!!! So it's the maximum possible to save very much power but still have more than enough performance in gaming on 4k@120Hz!
In the most games still I get my over 100 FPS and it's absolutely amazing ❤
same. UVed at 900mV and 2550Mhz + 141fps limiter (144fps - 3fps Headroom for possible framespikes).
Power Consumption went down from 439W peak (Borderlands 3 Benchmark, 1440p, Max settings + Reshade)
to 272W peak, with temps barely pass 45 degree Celcius, while maintaining constant 141fps.
@@BeatmasterAC excellent 😍✅
Still consider an undervolt if your trying to reuse say your old 750 watt power supply.Couple the undervolt with say an older 65 watt CPU and you should be able to do that ! Especially if you also consider those reported transient power spikes in the GPU's !
Can you speak more to the new data port from the PSU to GPU. Is it required? Any performance loss/gain with/without it? What is it’s purpose?
Yeah, this video should be removed terrible misinformation here
This video needs to be deleted. It's bad advice based on misinformation.