Rtx 3000 series undervolt discussion

More thorough testing is needed, but making the changes starting at +105 offset per your suggestion, I noticed the following in TimeSpy:

  • Prior to editing my curve, with GPU temp idling at 31C, my curve showed a peak clock speed of 2,085MHz from 1,137mV to 1,243mV
  • Average temp only went up only by 1C (64C vs 63C), but ambient temp in house is up by about 2C from the test I am comparing to (this week has been warmer than usual).
  • Peak clock speed is up by 30MHz (2,055MHz vs 2,025MHz)
  • Average clock speed is up by 51MHz (1,872MHz vs 1,821MHz)
  • Graphics score, not overall score, is up 167 points (19,787 vs 19,620), roughly .8%. I didn't want to use overall score since my CPU score has varied (but same clock speed as before).
  • Graphics Test 1 framerate is nearly identical (129.44 vs 129.45)
  • Graphics Test 2 framerate is up 1.78fps (113.08 vs 111.30), roughly 1.6%
My test today was on driver version 27.21.14.5730 and while the older benchmark was running on 27.21.14.5709. Power limit was left at 100%, Temp Limit at 83C, no added memory clock speeds (stock), and stock fan profile.
 
I just started messing with undervolt on the Aorus Xtreme 3080 - so far this is the best I've got that seems stable (running FurMark for awhile, haven't tested other things):

1605214136078.png


When you find something stable, do you try to bump up the clocks, or down the voltage, or what? When I tried .950 V it wasn't stable at these speeds. However, I get boosting to 2040 Mhz and temps seem to have dropped by almost 10 C.
 
Okay never mind, this one is better:
1605215074858.png


Sustained 2055-2070 Mhz, 63C max. I tried at 0.975 but it wasn't stable, bumping to 1 V seems to be good, and still dropped my temps from 71C at full load (was only maxing out at 2025 Mhz before as well).

Haven't messed with VRAM OC yet - is that necessary?
 
I just started messing with undervolt on the Aorus Xtreme 3080 - so far this is the best I've got that seems stable (running FurMark for awhile, haven't tested other things):

View attachment 298570

When you find something stable, do you try to bump up the clocks, or down the voltage, or what? When I tried .950 V it wasn't stable at these speeds. However, I get boosting to 2040 Mhz and temps seem to have dropped by almost 10 C.
using my guide as a baseline, bumping clocks is the same thing as downing voltage. Both occur. There is definitely a point where you are sacrificing stability if you are trying to set peak clock speed to too low a voltage, but if you increase incrementally by 15mhz, you usually dial it back before that point anyway, for instability at a different frequency. Also, furmark may or may not be the best test for stabilty; certain programs/games can cause the card to be power limited and not actually operate under the conditions that cause instability. Further, you want to test with an application that causes your card to fluctuate a bit in terms of clock speed, otherwise you are really only testing a single point on the curve.

You can OC Vram but it will take power away from the GPU core. If you are not power limited, it doesn't matter, but if you are, then you might hurt performance. It probably isn't worth it in real world use.
 
I only tested VRAM in SOTTR ranging from -250 MHz to +700 MHz, only saw a 1 FPS difference. Didn't do much testing though would be nice to get some more information on VRAM overclocks.
 
I went ahead did two other tests, one was dropping the power limit to 90% and the other I upped the power limit to 114% (temp limit to 90C) and added +500 to the memory. Both are using the same stock fan profile and same curve as before. The comparisons are to my prior test.


Test with Power Limit lowered to 90 % compared to original stock test:
  • Average temp is the same (63C vs 63C), remember that room is 2C warmer than when stock was tested.
  • Peak clock speed is up by 50MHz (2,070MHz vs 2,025MHz)
  • Average clock speed is down by 27MHz (1,794MHz vs 1,821MHz)
  • Graphics score, not overall score, is down 332 points (19,288 vs 19,620), roughly 1.7%.
  • Graphics Test 1 framerate is down by 2.51fps (126.94 vs 129.45), roughly 1.9%
  • Graphics Test 2 framerate is down 1.66fps (109.64 vs 111.30), roughly 1.5%
  • Peak power usage was around 315w
Test with Power Limit lowered to 90% compared to prior test:
  • Average temp decreased by 1C (63C vs 64C)
  • Peak clock speed increased by 15MHz (2,070MHz vs 2,055MHz)
  • Average clock speed is down by 78MHz (1,794MHz vs 1,872MHz)
  • Graphics score, not overall score, is down by 499 points (19,288 vs 19,787), roughly 2.5%.
  • Graphics Test 1 framerate is down 2.5fps (126.94 vs 129.44), roughly 1.9%
  • Graphics Test 2 framerate is up 3.44fps (109.64 vs 113.08), roughly 3%
  • Peak power usage was around 315w
Test with Power Limit at 114% and memory overclocked +500 compared to prior test:
  • Average temp increased by 3C (67C vs 64C)
  • Peak clock speed increased by 15MHz (2,070MHz vs 2,055MHz)
  • Average clock speed is up by 62MHz (1,934MHz vs 1,872MHz)
  • Graphics score, not overall score, is up 799 points (20,586 vs 19,787), roughly 4%.
  • Graphics Test 1 framerate is up 5.35fps (134.79 vs 129.44), roughly 4.1%
  • Graphics Test 2 framerate is up 4.47fps (117.55 vs 113.08), roughly 4%
  • Peak power usage was around 409w
That last test is putting me within spitting distance of my best benchmark I did earlier last week where I had similar settings except for instead of a clock curve, I had a flat +100MHz, had the core voltage set to +100%, and all fans at 100%. On that test my average temp was 54C, peak clock 2,115MHz, and average clock at 1,958MHz. That one had a graphics score of 20,765, only 179 points higher and within a 1% framerate difference.

So set to a 90% power limit (~315w), I am seeing about a 1.5-2% drop in performance over stock. Left at a 100% power limit (~350w), I am seeing about a 1-1.6% increase in performance over stock. Bumping the power limit to 114% (~400w) and memory +500MHz, I am seeing a 6.2-7.2% increase in performance over stock.
 
Is +500 on the memory a pretty safe bet, or did you tweak until you got that number?

EDIT: I see you are on a 3090 and I'm on a 3080, so I guess it might not equate over exactly.
 
Is +500 on the memory a pretty safe bet, or did you tweak until you got that number?

EDIT: I see you are on a 3090 and I'm on a 3080, so I guess it might not equate over exactly.
I was able to go all the way to +1500 without artifacting, not sure if I stressed it enough though. It didn't seem to have any effect on performance, I only tested it in Heaven.
 
Last edited:
Is +500 on the memory a pretty safe bet, or did you tweak until you got that number?

I was able to go all the way to +1500 without artifacting, not sure if I stressed it enough though.
But remember the memory is ECC and won't necessarily crash or fail, but will have an adverse effect on performance as well as create a lot of heat for the chips. The chips should have safety features built in to prevent temp damage, but again at the expense of performance.

The memory is rated at 21Gbps (21,000MHz, 1,312-1,313MHz x 16). The 3080 stock is clocked for 19Gbps (19,004MHz, ~1,187-1,188MHz x 16). The 3090 stock is clocked for 19.5Gbps (19,500MHz, 1219MHz x 16).

In Afterburner, a +500MHz memory overclock, if I am figuring it out correctly, is really a +63MHz overclock on the core memory clock. Afterburner's 500MHz reading is really a x 8 and not the x16, so the effective memory overclock is really 1,000 MHz after the x 16 calculation. This brings the memory speed to 20,504 MHz (or 20.5 Gbps). To hit the 21Gbps design limit (again, not calculating memory or power thresholds), you would have to overclock the memory to around 750-752MHz in Afterburner. Quite a few people on here seem to be able to hit 700MHz overclock on their memory on a 3090 but I don't feel comfortable doing that, mainly because my M.2 SSDs sit millimeters above the 3090 where it's memory resides and I don't want to heat those up more than necessary. With the 3080s (mainly the FE) being smaller and less ability to disperse heat, you may or may not have as much room to overclock the memory.

For me I think +500MHz is going to be as far as I push it until I put a water block on it just for heat's sake. I only set it to +500 for testing purposes here and will most likely either leave it at stock or maybe only bump it up +250MHz, which will have the memory running at an even 20Gbps, a .5Gbps bump over stock.
 
But remember the memory is ECC and won't necessarily crash or fail, but will have an adverse effect on performance as well as create a lot of heat for the chips. The chips should have safety features built in to prevent temp damage, but again at the expense of performance.

The memory is rated at 21Gbps (21,000MHz, 1,312-1,313MHz x 16). The 3080 stock is clocked for 19Gbps (19,004MHz, ~1,187-1,188MHz x 16). The 3090 stock is clocked for 19.5Gbps (19,500MHz, 1219MHz x 16).

In Afterburner, a +500MHz memory overclock, if I am figuring it out correctly, is really a +63MHz overclock on the core memory clock. Afterburner's 500MHz reading is really a x 8 and not the x16, so the effective memory overclock is really 1,000 MHz after the x 16 calculation. This brings the memory speed to 20,504 MHz (or 20.5 Gbps). To hit the 21Gbps design limit (again, not calculating memory or power thresholds), you would have to overclock the memory to around 750-752MHz in Afterburner. Quite a few people on here seem to be able to hit 700MHz overclock on their memory on a 3090 but I don't feel comfortable doing that, mainly because my M.2 SSDs sit millimeters above the 3090 where it's memory resides and I don't want to heat those up more than necessary. With the 3080s (mainly the FE) being smaller and less ability to disperse heat, you may or may not have as much room to overclock the memory.

For me I think +500MHz is going to be as far as I push it until I put a water block on it just for heat's sake. I only set it to +500 for testing purposes here and will most likely either leave it at stock or maybe only bump it up +250MHz, which will have the memory running at an even 20Gbps, a .5Gbps bump over stock.
I may try bumping to +250 to see if it's worthwhile. Thanks for the detailed reply.
 
Since the spec is 21,000MHz on the VRAM and it runs at 19,500MHz wouldn't that mean that the memory controller on the GPU is the bottleneck and not the actual memory itself?
 
Since the spec is 21,000MHz on the VRAM and it runs at 19,500MHz wouldn't that mean that the memory controller on the GPU is the bottleneck and not the actual memory itself?
If the bottleneck isn't ram temperature. Like a cpu might have a rated boost clock that's unsustainable with stock cooling. I was running +500 without problems but really didn't see the point other than going up a bit in benchmarks.
 
If the bottleneck isn't ram temperature. Like a cpu might have a rated boost clock that's unsustainable with stock cooling. I was running +500 without problems but really didn't see the point other than going up a bit in benchmarks.
I believe that the bottleneck, as you said, has more to do with temperature and power draw. These memory chips at full speed are pushing over 100C and I believe have a TjMax of around 110C and can be damaged if they reach or exceed 120C, if I read Igor Labs correctly. Of course, safety measures should be in place to prevent damage, but performance will suffer. That’s a lot of heat coming from these small chips, which will be a power draw and require ample cooling. It’s probably why they are clocked under their rated specs to run cooler and account for a wide range of installation scenarios where cooling may not be the best.

As such, I don’t think the memory controller is a limiting factor here.
 
If the bottleneck isn't ram temperature. Like a cpu might have a rated boost clock that's unsustainable with stock cooling. I was running +500 without problems but really didn't see the point other than going up a bit in benchmarks.
Yea so far I haven't seen a benefit to VRAM overclocking on my card other than benchmarks. Seems like a waste of time at this point.
 
I've done quite a bit of under-volt testing, and the sweet spot on my 3080 XC3 Ultra appears to be 1800Mhz @ .806v, completely stable. Pushing clocks harder generates a seemingly exponential increase in power for very little performance gain. Stock clocks settle at about 1845 @ 1.056v. Power goes from ~268w UV to ~400w stock.

I've mentioned it before, but I still find the whole thing baffling. It's as if at stock, the GPU is over-clocked like mad.
 
Bright memory infinite, you should be able to download it by clicking the text in the original post.
Thanks, will try it out.

In my experience, the game that can show instability where other games run fine is Quake 2 RTX. I’ve had to up the voltage/reduce clocks before it would be stable.
 
Hi,

i've just tried this curve on my Strix 3090,

https://zupimages.net/viewer.php?id=20/46/0edp.png

https://www.3dmark.com/compare/spy/15304171/spy/15274495

The score is not that much better than with 0.900 @ 1965hz and i got some perfcap reason pwr + pwr therm in gpu z even with maxed power limit

https://zupimages.net/viewer.php?id=20/46/qg3e.png

What does it mean for my card and overclock?
means you are power limited, and your card is getting hot. That curve you set is an extreme slope. Look at the first post step 6, a stepped curve should net better results.
 
Does upping the temp limit in Afterburner actually do anything if you're not hitting those temps? I only got about 70C max on my card without undervolting, now it's around 63C max. I have the boost set to 2070 but it still drops to 2055 for the most part, wondering if there is a way to keep it more stable.
 
Does upping the temp limit in Afterburner actually do anything if you're not hitting those temps? I only got about 70C max on my card without undervolting, now it's around 63C max. I have the boost set to 2070 but it still drops to 2055 for the most part, wondering if there is a way to keep it more stable.
Sounds like a power limit.
 
The
means you are power limited, and your card is getting hot. That curve you set is an extreme slope. Look at the first post step 6, a stepped curve should net better results.
The card wasn't that hot, 71c max
For the curve you mean it should be more regular? I have to set it manually? For the power limit is the limit coming from the gpu itself? Is there a way to max it? I see in gpuz that it hits 478 W so yeah I guess it's the reason but isn"t it very high for that kind of OC?
 
Last edited:
Here's what I did with my 3080 FE. I got crappy old 1080p monitors at 60hz, so 60fps is all I need. Does it in every game I've played at 4k DSR resolution. Main thing though is I have a Bitfinix Prodigy case and the 3080 gets hot unless I undervolt it. At stock it goes to 80c quick and throttles. I figure I didn't need those extra frames anyway since I got it vsync'ed to 60fps. Now the 3080 does 65 to 70 c and uses about 250 watts. I'm pretty happy with it. Now if only I can figure out a way to stick the 3080 in a NZXT H1 case and keep the temps down.


Capture.PNG
 
I noticed that when I set a custom curve in Afterburner, the core clock will sometimes go up one bin higher than the max clock speed specified e.g. 1965MHz even though the max set is 1950MHz@900mV in the Curve editor. I noticed this primarily happens when Folding. Does anybody know why?
 
I noticed that when I set a custom curve in Afterburner, the core clock will sometimes go up one bin higher than the max clock speed specified e.g. 1965MHz even though the max set is 1950MHz@900mV in the Curve editor. I noticed this primarily happens when Folding. Does anybody know why?
are you ever changing any settings in afterburner such as power limit, fan curve, etc? Any time you make a change and apply it in afterburner, the curve gets reapplied, and as I said in the first post of this thread, the curve is not based on absolute frequency, but rather OFFSET. Therefore, if your profile is saved with +180MHz offset at 900Mv which nets 1965MHz, and the card was warmer when you made that profile, reapplying it or making any other changes in afterburner and clicking apply will reapply your curve, but being cooler, the card may have a higher baseline, so 180MHz offset is a boost bin higher because the baseline curve was a boost bin higher because of lower temperature.
 
Hello,

Followed the guide in the OP and having ran Time Spy without any artifacts I scored lower than my "vanilla" run I had done a week ago.

Then I got rid of all unnecessary programs sitting at system tray and re-ran the test to see I broke the record for "AMD Ryzen 3 2200G + RTX 3080" combination (went even higher than the vanilla run).

1- Yeah my CPU is damn crappy but is that even possible? I did a COLD consecutive test, that's for sure and quitting those apps gave me another ~600 points in the re-run.

2- Also, what are we exactly doing by following the guide, trying to reach higher clocks at lower voltages so that we touch the upper threshold at a lower voltage?

3- Going over the recommended +105 Mhz increments by +15 Mhz each time strives for more performance or an even cooler curve? Which one? Both?

Thank you for tolerating my noobster questions and the awesome guide.
 
Hello,

Followed the guide in the OP and having ran Time Spy without any artifacts I scored lower than my "vanilla" run I had done a week ago.

Then I got rid of all unnecessary programs sitting at system tray and re-ran the test to see I broke the record for "AMD Ryzen 3 2200G + RTX 3080" combination (went even higher than the vanilla run).

1- Yeah my CPU is damn crappy but is that even possible? I did a COLD consecutive test, that's for sure and quitting those apps gave me another ~600 points in the re-run.

2- Also, what are we exactly doing by following the guide, trying to reach higher clocks at lower voltages so that we touch the upper threshold at a lower voltage?

3- Going over the recommended +105 Mhz increments by +15 Mhz each time strives for more performance or an even cooler curve? Which one? Both?

Thank you for tolerating my noobster questions and the awesome guide.
1. There can be run to run variations, and background programs absolutely affect the score to some degree. Also by running time spy cold, higher boosts will sustain for longer duration, giving you a higher average clock speed.

2. The card under operation will adjust what voltage it calls for based on temperature, power, etc. It doesn't explicitly call for clock speed, it calls for a voltage, and the curve you adjust changes the table where clock speed relates to voltage.

3. The whole guide is a combination of overclocking and undervolting, depending how you look at it. For a given voltage, we are overclocking. For a given speed, we are undervolting. Two ways of saying the same thing. Going over the 105MHz suggestion by 15MHz just further overclocks/undervolts, and will result in more performance across the voltage scale. In the OP example, the max speed on the default table was 2040MHz at 1.118v, and after setting the curve, it was 2040MHz at 1.000v, so that change alone, call it overclocking or undervolting (it is both), will result in more performance (as the card calls for 1.000v much more often than 1.118v) and less power draw due to the lower voltage.

Core clock is not exclusively the source of power draw on the card however, so if you leave power limit at 100%, there will still be cases you hit power limit. The core should run cooler in most cases, but really you want to set the power limit to something like 90% if you are trying to keep the card/case cool. Based on what the card is doing in terms of work and how much power it is using relative to power limit, the card will call for a voltage. As a hypothetical, let's say on our undervolt curve for a certain game, 100% power limit results in the card averaging 0.950v, and at 90% power limit set the card averages 0.900v. At this point you may say but at 0.950v the card does 1980MHz, vs the 90% power limit which takes me down to 0.900v and therefore only 1890MHz, why would I want to run the card slower?

The answer is relative. It would be true to say that after you set the undervolt curve, if you set a power limit ~90%, you are slowing down the card. BUT that is a choice you get to make as a user, the curve enables you to make this decision, because that 10% reduction in power is good for 32-35 watts of power and HEAT. And here comes the big reminder: if the card calls for 0.950v at 100% in the scenario above which results in 1980MHz on the custom curve, and at 90% calls for 0.900v which is 1890MHz, we can see that the difference in clock speed is 90MHz, right? BUT, to achieve this curve, we have increased these points on the curve for a given voltage by 105MHz. So if you compare 90% on the undervolt curve, which is 0.900v and 1890MHz, to the STOCK curve at 100%, which is 0.950v and (1980 - 105) 1875MHz, the undervolt curve at 90% power limit is 15MHz FASTER than the card would be stock, while producing 10% less power and heat.

That is a hypothetical again, but illustrates the point I think. Power draw is affected by things other than clock speed and voltage alone, but they definitely contribute. Any other questions I'm happy to answer.
 
The question "what are we exactly doing" is quite open to interpretation, or maybe preference. Mine was to find a good mix of performance and power while being stable. Slightly lower average clocks than stock, massively less power.

Seems to me that under-volting, whether using the OPs guide or another, is just a way to find a balance between performance, power, and temps. For my 3080, the balance point is an 1845mhz shelf @ .85v. It's stable, and keeps the power much lower than stock. With benchmarks, the under-volt is about 97% the performance of stock, and 95% the performance of the max mostly-stable over-clock I could get, while using 92 fewer watts. Stock on the 3080 seems to be a factory over-clock. Not going to get much more out of it without a bios that provides more power and some water cooling, but even then, I can't imagine it being worth much in actual games.
 
1. There can be run to run variations, and background programs absolutely affect the score to some degree. Also by running time spy cold, higher boosts will sustain for longer duration, giving you a higher average clock speed.

2. The card under operation will adjust what voltage it calls for based on temperature, power, etc. It doesn't explicitly call for clock speed, it calls for a voltage, and the curve you adjust changes the table where clock speed relates to voltage.

3. The whole guide is a combination of overclocking and undervolting, depending how you look at it. For a given voltage, we are overclocking. For a given speed, we are undervolting. Two ways of saying the same thing. Going over the 105MHz suggestion by 15MHz just further overclocks/undervolts, and will result in more performance across the voltage scale. In the OP example, the max speed on the default table was 2040MHz at 1.118v, and after setting the curve, it was 2040MHz at 1.000v, so that change alone, call it overclocking or undervolting (it is both), will result in more performance (as the card calls for 1.000v much more often than 1.118v) and less power draw due to the lower voltage.

Core clock is not exclusively the source of power draw on the card however, so if you leave power limit at 100%, there will still be cases you hit power limit. The core should run cooler in most cases, but really you want to set the power limit to something like 90% if you are trying to keep the card/case cool. Based on what the card is doing in terms of work and how much power it is using relative to power limit, the card will call for a voltage. As a hypothetical, let's say on our undervolt curve for a certain game, 100% power limit results in the card averaging 0.950v, and at 90% power limit set the card averages 0.900v. At this point you may say but at 0.950v the card does 1980MHz, vs the 90% power limit which takes me down to 0.900v and therefore only 1890MHz, why would I want to run the card slower?

The answer is relative. It would be true to say that after you set the undervolt curve, if you set a power limit ~90%, you are slowing down the card. BUT that is a choice you get to make as a user, the curve enables you to make this decision, because that 10% reduction in power is good for 32-35 watts of power and HEAT. And here comes the big reminder: if the card calls for 0.950v at 100% in the scenario above which results in 1980MHz on the custom curve, and at 90% calls for 0.900v which is 1890MHz, we can see that the difference in clock speed is 90MHz, right? BUT, to achieve this curve, we have increased these points on the curve for a given voltage by 105MHz. So if you compare 90% on the undervolt curve, which is 0.900v and 1890MHz, to the STOCK curve at 100%, which is 0.950v and (1980 - 105) 1875MHz, the undervolt curve at 90% power limit is 15MHz FASTER than the card would be stock, while producing 10% less power and heat.

That is a hypothetical again, but illustrates the point I think. Power draw is affected by things other than clock speed and voltage alone, but they definitely contribute. Any other questions I'm happy to answer.

Ah, thank you for the elaborate answer.

Just one more if you wouldn't mind please then:

As soon as Afterburner is closed, the curve is gone? What to do to keep it permanent and reset it to stock if I feel like it?
 
Ah, thank you for the elaborate answer.

Just one more if you wouldn't mind please then:

As soon as Afterburner is closed, the curve is gone? What to do to keep it permanent and reset it to stock if I feel like it?
Use the "Apply at Startup" option in afterburner once you have a stable curve.
 
The question "what are we exactly doing" is quite open to interpretation, or maybe preference. Mine was to find a good mix of performance and power while being stable. Slightly lower average clocks than stock, massively less power.

Seems to me that under-volting, whether using the OPs guide or another, is just a way to find a balance between performance, power, and temps. For my 3080, the balance point is an 1845mhz shelf @ .85v. It's stable, and keeps the power much lower than stock. With benchmarks, the under-volt is about 97% the performance of stock, and 95% the performance of the max mostly-stable over-clock I could get, while using 92 fewer watts. Stock on the 3080 seems to be a factory over-clock. Not going to get much more out of it without a bios that provides more power and some water cooling, but even then, I can't imagine it being worth much in actual games.
Regardless of methodology used, a power limit curve will always beat the flat curve in overall efficiency unless you are running the flat curve at 700mV then at best it can tie.
 
Regardless of methodology used, a power limit curve will always beat the flat curve in overall efficiency unless you are running the flat curve at 700mV then at best it can tie.
I think this point is difficult to see and explain. I know we discussed it in depth. Simply, a curve set like in the OP that tops out at the factory boost clock will perform better in many cases with a power limit cap vs a flat curve, even at the same power draw. This is because you are not always power limited, some game engines are more or less efficient, level of detail, resolution, all factor into how much power the card will draw. So in cases where you aren't power limited on a flat curve, the method in the OP will allow you to boost higher while still remaining within the constraints of the power limit of your choosing.
 
I wonder if nvidia has different or more stringent tests for stability. It would be a pretty big screwup to make the cards run 30-50W over their necessary power budget for no reason. Is there some kind of CUDA/compute stability test these days, something akin to Prime95 or IBT? If some of these curves are stable in games but unstable in compute, that would explain some things. Or maybe they really wanted the last 2-3% and decided it’s more important than 10+% higher power figures.
 
I wonder if nvidia has different or more stringent tests for stability. It would be a pretty big screwup to make the cards run 30-50W over their necessary power budget for no reason. Is there some kind of CUDA/compute stability test these days, something akin to Prime95 or IBT? If some of these curves are stable in games but unstable in compute, that would explain some things. Or maybe they really wanted the last 2-3% and decided it’s more important than 10+% higher power figures.
I think they knew AMD had competitive products this time around so they shipped the cards with more agressive clocks which in turn they had to loosen up the power levels for each frequency point because silicon quality varience.

There has always been 10-20% efficiency to be gained since atleast Maxwell, it's just more this time because GDDR6X.
 
are you ever changing any settings in afterburner such as power limit, fan curve, etc? Any time you make a change and apply it in afterburner, the curve gets reapplied, and as I said in the first post of this thread, the curve is not based on absolute frequency, but rather OFFSET. Therefore, if your profile is saved with +180MHz offset at 900Mv which nets 1965MHz, and the card was warmer when you made that profile, reapplying it or making any other changes in afterburner and clicking apply will reapply your curve, but being cooler, the card may have a higher baseline, so 180MHz offset is a boost bin higher because the baseline curve was a boost bin higher because of lower temperature.
Yeah, it looks like it was a matter of the boost bin being a step higher. I changed the offset and now get the maximum frequency to land where I want it. Thanks.

I'm noticing that the curve I set gets reset after a while, though. As in, it goes back to the stock curve instead of the specific frequency/voltage offsets I put in. I'm usually folding when not gaming (which has helped identify good long term stability incidentally) but when I come back, inevitably the curve is back to normal. Looking at the frequency editor shows nothing has changed. I am using Afterburner 4.6.3 beta 3. Anybody else seen something similar?
 
Use the "Apply at Startup" option in afterburner once you have a stable curve.
Thank you for this.

I've been seeing "power limit" being mentioned a lot. Does this need additional attention or is it a concept we have already meddled with by undervolting/oc'ing ?
 
Thank you for this.

I've been seeing "power limit" being mentioned a lot. Does this need additional attention or is it a concept we have already meddled with by undervolting/oc'ing ?
Its part of the guide in the op. Once you set your curve, youll want to set a power limit. I suggest between 80 and 90 but you should try it and see where you are happy with performance and Temps. At 90 you should basically be as fast as stock while running cooler. If you want even less power draw you can do 80 and sacrifice a few percent of performance for much lower power.
 
Yeah, it looks like it was a matter of the boost bin being a step higher. I changed the offset and now get the maximum frequency to land where I want it. Thanks.

I'm noticing that the curve I set gets reset after a while, though. As in, it goes back to the stock curve instead of the specific frequency/voltage offsets I put in. I'm usually folding when not gaming (which has helped identify good long term stability incidentally) but when I come back, inevitably the curve is back to normal. Looking at the frequency editor shows nothing has changed. I am using Afterburner 4.6.3 beta 3. Anybody else seen something similar?
I haven't noticed going back to the stock curve, and I have done a mix of gaming, deep learning, and compute. Latest non beta.
 
Back
Top