NVIDIA GeForce RTX 4080 Super Founders Edition

[chrcoluk]

I have now ran 3dmark tests, just to make sure its stable at stock and it is. The tests were ran uncapped frame rate with 100% limit, and the GPU peaked at less than 50C with case panel installed. Under 50C with over 300W pumped through it and 99% utilisation.

 

[chrcoluk]

Looking at my 4080 super frequency curve right now, it tops out at 2850mhz with 1240mv, losing a mere 30mhz it is 1160mv quite a difference, at 2490mhz is where the curve starts to flatten and lose efficiency, that is at 985mv where I think a card released 10 years ago would be capped at. So 2490mhz probably should have been stock clock.

Changing my frequency to cap at 2820mhz is a clear no brainer, I think I will make a second curve capping at 2490mhz used for games where the extra clocks are not needed, or maybe 2625mhz at 1040mv might be better, and also a curve that caps at 1418mhz 700mv for legacy games.

On the below graph, 2625 is the last one before blocks start being side by side (meaning larger voltage jumps needed for clock step), and of course 1418 is right at the base.

May or may not undervolt on top of that.

 

[Lew Zealand]

Looking at my 4080 super frequency curve right now, it tops out at 2850mhz with 1240mv, losing a mere 30mhz it is 1160mv quite a difference, at 2490mhz is where the curve starts to flatten and lose efficiency, that is at 985mv where I think a card released 10 years ago would be capped at. So 2490mhz probably should have been stock clock.

Changing my frequency to cap at 2820mhz is a clear no brainer, I think I will make a second curve capping at 2490mhz used for games where the extra clocks are not needed, or maybe 2625mhz at 1040mv might be better, and also a curve that caps at 1418mhz 700mv for legacy games.

On the below graph, 2625 is the last one before blocks start being side by side (meaning larger voltage jumps needed for clock step), and of course 1418 is right at the base.

May or may not undervolt on top of that.

View attachment 364748

The 4080 will not run at more than 1.075v unless you manually unlock voltage control, so in that graph you'll top out at 2745 MHz as seen here in the review:

While you will get additional efficiency by running at a lower voltage and clock speed, the real efficiency comes from overclocking to about +200 MHz and then flattening the top of that graph.

IMO the 2 best places to do that are generally at 0.95 - 1.00 V (for efficiency with stock performance) and then the lowest voltage your GPU runs at (for max efficiency and still good performance). Again from the review the 4080 Super runs at a minimum of 0.905 V (VSync line):

Looking at your graph:

0.905 V is at about 2200 MHz, or 2400 MHz when you apply the +200 MHz overclock
1.000 V is at about 2550 MHz, or 2745 MHz when you apply the +200 MHz overclock.

Note how that after the OC, the 1.00 V clock is also the card's max typical clock speed. But you'll be using about 20% less power at 1.00 V

I do the same on my 4060 Ti (160W GPU) but the numbers are different as it's minimum running voltage is 0.86v, so I typically run at:

2400 MHz 0.86 V, 105W max
2715 MHz 0.95 V, 125W max

Also OCing your VRAM is essentially free performance as it adds a watt or 3 to the total and can deliver 5% more fps, scene-dependent. Mine is OC'd from 2250 to 2600 MHz. My favorite VRAM tester is looped Time Spy Graphics test 2. While there may be a better VRAM test, I have found on 20+ GPUs that if it's stable with no artifacts in that test, it's been 100% stable in every game I play.

 

[chrcoluk]

The 4080 will not run at more than 1.075v unless you manually unlock voltage control, so in that graph you'll top out at 2745 MHz as seen here in the review:

While you will get additional efficiency by running at a lower voltage and clock speed, the real efficiency comes from overclocking to about +200 MHz and then flattening the top of that graph.

IMO the 2 best places to do that are generally at 0.95 - 1.00 V (for efficiency with stock performance) and then the lowest voltage your GPU runs at (for max efficiency and still good performance). Again from the review the 4080 Super runs at a minimum of 0.905 V (VSync line):

View attachment 364793

Looking at your graph:

0.905 V is at about 2200 MHz, or 2400 MHz when you apply the +200 MHz overclock
1.000 V is at about 2550 MHz, or 2745 MHz when you apply the +200 MHz overclock.

Note how that 1.00 V clock is also the card's max typical clock speed. But you'll be using about 20% less power at 1.00 V

I do the same on my 4060 Ti (160W GPU) but the numbers are different as it's minimum running voltage is 0.86v, so I typically run at:

2400 MHz 0.86 V, 105W max
2715 MHz 0.95 V, 125W max

Also OCing your VRAM is essentially free performance as it adds a watt or 3 to the total and can deliver 5% more fps, scene-dependent. Mine is OC'd from 2250 to 2600 MHz. My favorite VRAM tester is Time Spy Graphics test 2. While there may be a better VRAM test, I have found on 20+ GPUs that if it's stable with no artifacts in that test, it's been 100% stable in every game I play.

Makes sense thanks, I havent seen anything higher than 2700mhz in testing. Wasnt planning on overclocking core clocks, as silicon lottery (but maybe could experiment combined with the custom curves), although might do memory as long as it doesnt have a significant impact on temps or wattage, apparently on the 4080 super the memory chips are spec'd higher than what Nvidia run them at.

Ok I have done the new curves I will use for now, using this method, I was less aggressive on the clock offset. I noticed cant really do what I did on the 3080 because these cards wont go below 0.900v, so possibly might be less efficient in very old games, wattage saving on the undervolted curves is ok, I did also bump the memory clock as Nvidia underclocked it, but in everything I tested it didnt do anything outside of margin of error, but since it seems to barely impact power consumption I left it in case I run a game that appreciates the higher memory bandwidth.