3090 Undervolting

[wunkley]

I’ll give Quake II a go today. Ive checked with Control, Cyberpunk, Metro Exodus with no dlss with high ultra raytracing. I find that I can pass benchmarks but as soon as I try the games with raytracing it was crashing with lower voltages, but with this it’s fine. I mean I’m never going to play these games without dlss anyway but I’d rather be fully stable.

I’ve seen some people achieve over 2000mhz. I couldn’t get anywhere near that as I hit the power limit straight away, even in benchmarks. Not that I would go that high, but in terms of golden chips, I have seen some crazy numbers claimed.

it is the same with overclocking my cpu. I can pass Linpack extreme, prime 95, real bench etc for hours, but I get whea errors after ten minutes of Apex Legends or Cod on certain modes. Not that I play those games, but I use them for testing stability. They are much better than any benchmark in my opinion. Real world testing has shown much more effective with any stability for me.

What settings are you using for Quake II, if you don’t mind?

If Control works ok then you're probably good with those curve settings. For Quake II rtx I'm just using the default settings it loads with.

Overclocking to the max can be a bit of an obsession for not that much real world gain most of the time in actual games. I'd rather run cooler and quieter, as long as the performance is good enough for my wants.

 

[Steve Hel]

If Control works ok then you're probably good with those curve settings. For Quake II rtx I'm just using the default settings it loads with.

Overclocking to the max can be a bit of an obsession for not that much real world gain most of the time in actual games. I'd rather run cooler and quieter, as long as the performance is good enough for my wants.

Im struggling with Quake II. Literally can’t get this stable. I see what you mean now. I think the only way is to drop it down to 1750 and hopefully I can manage it.

 

[gpuerrilla]

Overclocking to the max can be a bit of an obsession for not that much real world gain most of the time in actual games. I'd rather run cooler and quieter, as long as the performance is good enough for my wants.

Agree, from the 90s to the past couple of years I now aim for an underclocked tweaked profile that has less heat and noise for 98% of the performance.

 

[Joxeon]

Agree, from the 90s to the past couple of years I now aim for an underclocked tweaked profile that has less heat and noise for 98% of the performance.

Yeah this this is definetly the way to go with ampere since the last 20% power only provides around 5% extra performance as nvidia cranked these up to make sure they beat out AMD.

 

[newuser]

Overclocking to the max can be a bit of an obsession for not that much real world gain most of the time in actual games.

Balance in all things, the unbroken circle of Zerthimon. God I'm old...

I'm aiming for safe temps at acceptable noise levels. I often run my cpu flat out for hours while rendering but it runs at 70c with a mild all core oc without the cooler sounding like a dustbuster. Trying to do the same with the gpu.

 

[Steve Hel]

To get quake II stable, I need to drop to 1785, 806, or I can go to 1740, 793. Just with you mentioning that anything over 0.83 volts isn’t going to save me any power in full raytracing games, I thought better safe than sorry.

I’m glad I tried Quake II, it’s absolutely brutal. I honestly thought I was fine with my 1905, 856v how wrong was I lol.

It saves a huge amount of power and heat in normal non brutal raytracing games, so happy days.

 

[ToOo]

Sorry for the slight necro but the information in this is golden especially from @wunkley , thanks very much for providing some real data!

I found what I had thought was a pretty good guide for undervolting at optimum tech, but I couldn't replicate the results at all and i've been banging my head againt it all weekend.


Using the information in this thread as a reference my own testing with a 3090fe shows at 0.806v and 1785mhz Quake 2 RTX will eventually climb to just over 380W power consumption if you increase the power limit to maximum and let it run:

Time:-----GPU Temp:-----Peak reported GPU Power:
0s------50C-----------340W
35s-----60C----------355W
70s-----65C----------365W
110s----70C----------370W
200s----75C----------380W (reached thermal equilibium at this point)

Test was run at 1440p with high global illumination and standing still in a room with no movement or enemies, the theory being that the load should have been completely constant. There seems to be a clear trend for increased power consumption as the GPU temperature increases with even relatively minor thermal differences incrementing the power consumption significantly.

Testing something more trivial like unigen Haven 4 showed similar results, starting with peaks at about 270W and hitting just over 305W by the time it has reached thermal equilibrium.

I've already re-padded and repasted my card but i'm going to try some better paste during the week and see if I can make any change to these numbers. I'm not willing to try liquid metal so i dont expect a significant difference.

It would be really interesting if some others could try and replicate these results.

 

[Radox-0]

Sorry for the slight necro but the information in this is golden especially from @wunkley , thanks very much for providing some real data!

I found what I had thought was a pretty good guide for undervolting at optimum tech, but I couldn't replicate the results at all and i've been banging my head againt it all weekend.

Using the information in this thread as a reference my own testing with a 3090fe shows at 0.806v and 1785mhz Quake 2 RTX will eventually climb to just over 380W power consumption if you increase the power limit to maximum and let it run:

Time:-----GPU Temp:-----Peak reported GPU Power:
0s------50C-----------340W
35s-----60C----------355W
70s-----65C----------365W
110s----70C----------370W
200s----75C----------380W (reached thermal equilibium at this point)

Test was run at 1440p with high global illumination and standing still in a room with no movement or enemies, the theory being that the load should have been completely constant. There seems to be a clear trend for increased power consumption as the GPU temperature increases with even relatively minor thermal differences incrementing the power consumption significantly.

Testing something more trivial like unigen Haven 4 showed similar results, starting with peaks at about 270W and hitting just over 305W by the time it has reached thermal equilibrium.

I've already re-padded and repasted my card but i'm going to try some better paste during the week and see if I can make any change to these numbers. I'm not willing to try liquid metal so i dont expect a significant difference.

It would be really interesting if some others could try and replicate these results.

Yes, I have been noticing something similar for a while now, it became more pronounced when I moved to under water which shaved nearly 30 degrees off under load. Another benefit of this was I can actually stabilise higher clock speeds with lower voltages so it creates a cycle almost, of colder temps needing less voltage to achieve a certain clock speed which in turn needs less power which likely stabilises things with less power / voltage leak occurring in the core. However as you also describe even on air, as the heatsink warms up you can notice the power draw increase.

Even on water there is an effect. For my AC Valhalla profile I use 1.0v @ 2100Mhz + 1000 on the memory. When my PC starts up initially the benchmark run in the game with GPU at 30 degrees, power use is around 285w during the run. As GPU warms and averages 35 degrees during run, this goes to 290w and at 40-42 degrees which is where my loop normally stabilises the GPU it consumes about 300w.

Also as I initially mentioned, colder temps do also allow you to use less voltage. My Cyberpunk 2077 overclock was initially 0.9v @ 2040Mhz. Worked fine for days then noticed it kept crashing. I literally traced it down to the temp of my loop. Initially when I set up that 0.9v profile, the room was cold and fans on loop running quicker so GPU averaged 35 degrees. As I have fine tuned the loop and lowered the Fan RPM the the GPU tops out around 40 degrees the clock speed I now set is 2010Mhz, so two steps lower.

Its pretty interesting overall seeing the temp change impact of these cards, at least from what I can see.

 

[chrcoluk]

test port royale and timespy, if you at the limit, then go up 2 voltage levels, and will probably be stable.

850mv wont be possible at 1900+ on all cards, my ampere needs 874mv and I run at 882mv day to day.

Not sure if voltages required for 3090s are higher than 3080 or not.

I have seen 892mv banded about on reddit, 850mv actually seems quite aggressive, but the best cards can handle it.