RTX 4070 - efficiency & undervolting

[THU31]

Undervolting is the new overclocking. I've been using it since the Pascal era, and with Ampere it proved to be an incredible way of reducing the gigantic power consumption, while retaining almost all of the performance.

I decided to replace my MSI RTX 3080 GamingX Trio with the MSI RTX 4070 Ventus 3X.
First thing I wanted to test was how effective undervolting was with such a relatively low-power AD104 card. I also wanted to compare it to my undervolted 3080, since these cards offer pretty much identical performance.

Here are the results. All testing was done with "Prefer max performance" power management.

3DMark TimeSpy (1440p, no Vsync)

4070 - stock settings (2670-2760 MHz @ 1.01-1.06 V, average clock ~2715 MHz)
Graphics score - 17309
Graphics test 1 - 112.93 FPS
Graphics test 2 - 99.15 FPS
Power draw - 197-202 W (constantly in power limit)

4070 - fixed 2805 MHz @ 1.0 V
Graphics score - 17457
Graphics test 1 - 113.95 FPS
Graphics test 2 - 99.95 FPS
Power draw - 177-200 W, average ~190 W (hits the power limit a few times, the clock drops to 2775-2790 for a moment)

4070 - fixed 2610 MHz @ 0.91 V
Graphics score - 16658
Graphics test 1 - 108.73 FPS
Graphics test 2 - 95.38 FPS
Power draw - 141-158 W, average ~150 W

3080 - fixed 1800 MHz @ 0.8 V
Graphics score - 16902
Graphics test 1 - 110.38 FPS
Graphics test 2 - 96.74 FPS
Power draw - 252-281 W, average ~270 W


Forza Horizon 5 (in-game benchmark, capped 4K60 with Vsync, Ultra settings with TAA)
Game is known to run extremely well on Ada Lovelace. I assume the benchmark estimates the framerate values when running with a capped framerate.

4070 - stock settings (constant 2805 MHz @ 1.10 V)
Average FPS - 94
Minimum FPS - 81
Power draw - 126-181 W, average ~155 W

4070 - fixed 2805 MHz @ 1.0 V
Average FPS - 91
Minimum FPS - 79
Power draw - 104-152 W, average ~130 W

4070 - fixed 2610 MHz @ 0.91 V
Average FPS - 88
Minimum FPS - 76
Power draw - 90-128 W, average ~115 W

3080 - fixed 1800 MHz @ 0.8 V
Average FPS - 83
Minimum FPS - 73
Power draw - 172-235 W, average ~200 W


Destiny 2 (30-minute mission with a lot of chaos, capped 4K60 with Vsync)

4070 - fixed 2610 MHz @ 0.91 V
Power draw - 95-140 W, average ~110 W

3080 - fixed 1800 MHz @ 0.8 V
Power draw - 190-260 W, average ~210 W


Fallout 3 (3-minute run through the open world, capped 4K60 with Vsync, GPU usage 20-35% on both cards)

4070 - fixed 2505 MHz @ 0.91 V (default boost clock, won't go any higher with such low GPU usage)
Power draw - 50-65 W, average ~55 W

3080 - fixed 1800 MHz @ 0.8 V
Power draw - 94-141 W, average ~120 W


TessMark (3 minutes of demo mode)

4070 - fixed 2610 MHz @ 0.91 V
Power draw - 115-131 W, average ~122 W

3080 - fixed 1800 MHz @ 0.8 V
Power draw - 190-225 W, average ~210 W


MPC-HC with madVR (540p upscaled to 4K with Jinc AR)
Video playback requires "Prefer max performance" for perfect results with no stutter or dropped frames, which are caused by the GPU constantly switching power states with Normal/Adaptive management.

4070 - fixed 2610 MHz @ 0.91 V
Power draw - 40-47 W, average ~44 W

3080 - fixed 1800 MHz @ 0.8 V
Power draw - 105-110 W, average ~107 W


The RTX 4070 is extremely efficient even at stock settings, but it will hit the power limit in most scenarios where the framerate is uncapped.
A standard undervolt of 2805 MHz @ 1.0 V can reduce the average power draw by 5-15% while retaining stock performance.
An extreme undervolt of 2610 MHz @ 0.91 V can reduce the average power draw by 25% while retaining 95% of stock performance (or identical performance with a capped framerate)
Compared to an extremely undervolted RTX 3080, an extremely undervolted RTX 4070 offers a 40-50% reduction in power draw across the board, as much as 120 W in my testing.
I did not test if the clocks can go any higher at those voltages, I based them on other undervolting results. From what I saw with the 4070 Ti undervolting, AD104 cards can achieve higher clocks at the same voltages.
Idle voltage is 0.89 V (power draw is 13-15 W without any link state management). The minimum voltage in boost mode is 0.91 V. I wonder if this is a limitation of the TSMC 4N node. Ampere cards can go as low as ~0.725 V.

This test is not supposed to convince anyone that the RTX 4070 is a great value card. It's just meant to showcase the efficiency of the Ada Lovelace architecture, especially compared to Ampere.

 

[Vayra86]

Cool tests!

 

[Dristun]

Interesting! Wonder if other cards that have 8- and 10-phase designs for voltage can do even better.

 

[oxrufiioxo]

Very neat the whole ADA family can be ridiculously power efficient I can get my 4090 down to around 300-340W while losing hardly any performance around 5-6% at most.


This is honestly the 4070s most desirable trait how much less power it uses vs similarly performing last generation cards. Also and maybe this is a bigger reflection of how overpriced it is you can easily buy it unlike at the start of last generation.

 

[Buchling]

I just got the INNO3D RTX 4070 X3 OC and used your settings as a reference. Im currently running my card with 2820 MHz at 1.0 Volts. Power draw is ~180 Watts and the memory is overclocked to 11,000 MHz. The maximum temperature is around 65°C and the fans stay at around 40%. Pretty happy with these results so thank you for this thread. The current clock is not 100% stable though.

 

[NucLeaR]

not bad tests, but it would be nice to check how it behaves at a voltage of 0.95v. I dare to assume that the optimal operating mode will be achieved at this voltage.

 

[Buchling]

I am currently running my card at 2720 MHz and 0.95 volts. Power draw is around 160 watts, the temperature stays under 60 degrees. There is almost no performance loss to standard settings.

 

[Mussels]

I love the insane difference undervolting made on my 3090, i've got plans to do it with my 3070 as well.
375W down to 200-230W depending on the game, from a 10% drop in clocks (erratic up-to 2GHz to constant 1800Mhz)

You broke your own testing by forcing this

"All testing was done with "Prefer max performance" power management."

You forced your 3080 to use more power than my 3090


Stock runs between 1750 and 2000, depending on the load on this GPU.

Minimum voltage setup, 0.737v 1600Mhz

My GPU will often downclock below the undervolted clock speed and save even more power - espcially with an FPS cap - and you've disabled that feature.


I've lost my clocks in afterburner somehow, i'll try and re-test at 1700 and 1800 when i find the text file i recorded my voltages in.


Edit: found em
0. As above, 1600 @ 737 is 190W
1. 1675 737mv - 240W
2. 1750MHz - 750mv 253W
3. 1800 - 775 269W
4. 1830 - 815mv (didnt record wattage)
5.1900 837mv (360W)

 

[THU31]

My GPU will often downclock below the undervolted clock speed and save even more power - espcially with an FPS cap - and you've disabled that feature.

I tested this and it has a negative impact on frametimes and can cause microstutters. And once the GPU utilization is low enough, it will actually switch the power state down from P0 and that causes a huge hitch. And sometimes it can go back and forth between different power states, especially in older less demanding games.

Personally I prefer maximum power saving at maximum performance, for me it's the right balance.

It's the same with CPU power management. Variable clocks and C-states have a very negative impact on frametimes. That's why I prefer to manage it myself, set a fixed speed according to what game I'm playing. I like power saving as long as it doesn't compromise my fluidity. Many games stutter enough as it is, without adding extra stutters from power management.

 

[Bojji]

I am currently running my card at 2720 MHz and 0.95 volts. Power draw is around 160 watts, the temperature stays under 60 degrees. There is almost no performance loss to standard settings.

I'm running mine at 2700Mhz with 0.94v, it performs almost exactly like default setings in terms of performance.

I also have profile with 2800MHz/0.985v (probably needs 0.99 to be perfectly stable) that performs better than stock.

Overal it's very efficient GPU.

 

[Lew Zealand]

not bad tests, but it would be nice to check how it behaves at a voltage of 0.95v. I dare to assume that the optimal operating mode will be achieved at this voltage.

Agreed. My Pascal, Turing and RDNA2 cards seem to converge near 0.95v being the inflection point where above this the clocks & FPS increases flatten out with large power increases and below this the power and clocks/FPS drop more linearly. Could be that Ada (and RDNA3) has a different voltage point for that though, would be nice to know.

 

[tabascosauz]

I tested this and it has a negative impact on frametimes and can cause microstutters. And once the GPU utilization is low enough, it will actually switch the power state down from P0 and that causes a huge hitch. And sometimes it can go back and forth between different power states, especially in older less demanding games.

Personally I prefer maximum power saving at maximum performance, for me it's the right balance.

It's the same with CPU power management. Variable clocks and C-states have a very negative impact on frametimes. That's why I prefer to manage it myself, set a fixed speed according to what game I'm playing. I like power saving as long as it doesn't compromise my fluidity. Many games stutter enough as it is, without adding extra stutters from power management.

I guess this is highly situational from setup to setup. On Turing and Pascal I've stayed away from Prefer Maximum Performance because idle power is much higher, but it seems to make no difference on 4070 Ti.

Funnily enough the downclocking was what made me walk away from RDNA3 because on the Radeon side there is no user control over it. But my 4070 Ti is smart enough to keep VRAM clock high on old/light games to prevent frametime problems, even on Optimal Power - effect of low core clock is usually negligible, it's the VRAM clock dropping that causes problems.

 

[blackie]

Undervolting is the new overclocking. I've been using it since the Pascal era, and with Ampere it proved to be an incredible way of reducing the gigantic power consumption, while retaining almost all of the performance.

I decided to replace my MSI RTX 3080 GamingX Trio with the MSI RTX 4070 Ventus 3X.
First thing I wanted to test was how effective undervolting was with such a relatively low-power AD104 card. I also wanted to compare it to my undervolted 3080, since these cards offer pretty much identical performance.

Here are the results. All testing was done with "Prefer max performance" power management.

View attachment 291919

Thank you, your results helped me a lot. Tried to undervolt Asus Dual 4070 before but was not stable.
Glad to confirm both your profiles are working perfectly, use the power saver profile 2610@0.91V as my daily driver.
4070 is finally at efficiency level I had hoped

 

[Mussels]

Tried to undervolt Asus Dual 4070 before but was not stable.

It's important to note different loads have different stability - something that lasted hours in a DX11 test for me at 100% GPU usage crashed in minutes in a DX12 title with fluctuating usage, it's always best to document what seems stable and then raise volts/lower clocks slightly after each error. I've seen oddities like a flat line in afterburner still varying (Internal multiplier/divider used to generate the speeds?) so you could get a 25-50MHz variance from what should have been a static value.

 

[tugrul_SIMD]

Ventus 2x oc edition rtx4070: 925mV, 2640 MHz, 140Watts.

 

[wolf]

Undervolting is the new overclocking.

I wonder how I missed this post, I love that quote and for a very long time it was in my Sig too.

Also been doing it since pascal and it's awesome that despite the massive efficiency gains over ampere, there are still mad gains to be had with undervolting.

My napkin math indicates we'll be able to get the GPU horsepower of a 2080Ti / 3070 / 6700XT / 6800 in a low profile slot powered card if Nvidia wanted. Well theres the awesomely ludicrous Ada A4000 20GB, but the price is yikes, I want say a 4060/ti in that form factor.

 

[Mussels]

I wonder how I missed this post, I love that quote and for a very long time it was in my Sig too.

Also been doing it since pascal and it's awesome that despite the massive efficiency gains over ampere, there are still mad gains to be had with undervolting.

My napkin math indicates we'll be able to get the GPU horsepower of a 2080Ti / 3070 / 6700XT / 6800 in a low profile slot powered card if Nvidia wanted. Well theres the awesomely ludicrous Ada A4000 20GB, but the price is yikes, I want say a 4060/ti in that form factor.

My 3070 can run at ~120W, the performance of a 2080Ti at the wattage of a 1660 Super

My brother just got a new laptop and the same is on the todo list there, curve optimiser on the CPU and a PPT limit, followed by afterburner on the 3050 6GB with it - should be able to nearly halve the stock wattages

 

[ratirt]

Nice UnderOC bro and an awesome achievement. The undervolting is a great thing to be honest and it does miracles. I see your statistic and it is impressive how much you can achieve just by lowering some voltage and a some tweaking. Power consumption cut by a lot.

 

[squallypo]

@THU31 could you explain how did you applied the undervolt/overclock method that you used?
https://www.reddit.com/r/nvidia/comments/wghb9d i ask because of this thread.

 

[THU31]

@THU31 could you explain how did you applied the undervolt/overclock method that you used?

Method 1 from that post. It's the best way to achieve a 100% fixed clock speed for gaming with a capped framerate.

 

[squallypo]

Method 1 from that post. It's the best way to achieve a 100% fixed clock speed for gaming with a capped framerate.

would you say method 2 is also fine or not recommended?

 

[Vayra86]

I wonder how I missed this post, I love that quote and for a very long time it was in my Sig too.

Also been doing it since pascal and it's awesome that despite the massive efficiency gains over ampere, there are still mad gains to be had with undervolting.

My napkin math indicates we'll be able to get the GPU horsepower of a 2080Ti / 3070 / 6700XT / 6800 in a low profile slot powered card if Nvidia wanted. Well theres the awesomely ludicrous Ada A4000 20GB, but the price is yikes, I want say a 4060/ti in that form factor.

If Ada was going in 1 slot form factors every other PC in my house would have had one now. That is, if they're not 600 bucks

They're bloody stupid & lazy

'We need moar' is all they (we?) think of. Complete tunnelvision, despite the fact even a toaster can run most games at a low res now. It has always baffled me how ever faster GPUs still manage to be so stagnant in practice, and even getting worse ever since RT was stuck on. Bigger, more power hungry... and we're still looking at utter shite FPS in comparison. None of this is in our benefit really; didn't we all just want to play games and have fun?

Look at those crowds ooh-aah ing the next 30% hard perf increase every time... and what do they do with it? Run 4K, even though half the pixels is faked. Buy a 240hz monitor so they can whine about not having good steady frametimes anymore (muh CS GO fps nuuu!) Its completely silly
And I'm not immune either, myself. But yea. Damn

 

[tommesfps]

Yeah, ok, good study, but hey, you bought a 600$ graphics card, you are supposed to squeez every single fps out of it according to major hardware review outlets all over the world. Please put in 1000Watt psu again to play 4K uncapped ultra extreme settings for potential best visual experiance, and remember 89fps is better than 77fps. KKTHXBYE over and out...

 

[Onyx Turbine]

Anyone has ideal settings for a 3070? What are thoughts about 3070 undervolting compared to a 4070 with as purpose ocasionally game at high settings (not ultra) in 1440p.
As I find the price gap between 3070 and 4070 significant, is dls 3 that much better than dls 2, better said is a 3070 not future proof enough?

 

[Assimilator]

This is what node shrinks are SUPPOSED to achieve: much lower power consumption at almost the same performance levels. But because the vendors have collectively lost their minds, everything is pushed way past its limits, forever. It honestly disgusts me.