Xatos1337
My graphics score out of the box in 3D Mark was 19,498. I was thinking about letting her rip and seeing how far above 20K she can go. Anyone have any safe OC numbers yet that I can use as a baseline? I don't really ever do any OC'ing but this card feels like she's being held back and I want to let breathe for a minute. 

Nvidia has these chips so power limited that there really is no "unsafe" OC settings. You can safely crank the power limit all the way up and tinker from there. I'd try for 2000Mhz and see if it crashes. If not nudge it up a bit from there. The worst you can do is crash the driver. The card will never pull enough power to break it.

justin_43

Xatos1337
My graphics score out of the box in 3D Mark was 19,498. I was thinking about letting her rip and seeing how far above 20K she can go. Anyone have any safe OC numbers yet that I can use as a baseline? I don't really ever do any OC'ing but this card feels like she's being held back and I want to let breathe for a minute. 

Nvidia has these chips so power limited that there really is no "unsafe" OC settings. You can safely crank the power limit all the way up and tinker from there. I'd try for 2000Mhz and see if it crashes. If not nudge it up a bit from there. The worst you can do is crash the driver. The card will never pull enough power to break it.

Sounds like solid advice. Thanks!
 

talon951
After some fiddling I got +165 core and +1100 to run in 3dmark PR. Probably not completely stable. For some reason it didn't save that run online though. It was 14350 or so. I'll run it again tonight when I get home.

I'd say start at +135 core and +750 mem with power maxed at 113%. If that completes, move up from there in small increments.

This is an earlier run at 135/1000.

https://www.3dmark.com/pr/1072918

Awesome thanks for the input. I'll try out some OC'ing later and report back!

https://www.3dmark.com/3dm/62580515?
165 core / 1000 memory
https://www.3dmark.com/3dm/62580687?
150 core / 1000 memory
 
Pretty damn happy, card purrs along at 55c, 3080Ti FTW3 Hybrid with push/pull set up on the rad.

• 150 core
• 1,000 mem
• https://www.3dmark.com/spy/20817155

 
Timespy GPU Score: 20,446
 

Must not be anyone running Tim Spy Extreme.  4th place graphics score on my old 4790k.  2000 cpu score. LOL
 
https://www.3dmark.com/spy/20817728

https://www.3dmark.com/3dm/62585446?
 
GPU Score 20,875

So for an all gaming overlock it seems my card can only boost to 2010 at Max. 2025 and higher it seems to crash games after an hour. Memory is +1000 though.  my 3080 got to about 2080 MHz core, but I think most 3090/Ti don’t go that high long term

So far +135 core and +1100 mem seem to be game stable.  No crashes in CP2077 and it passed the 3dmark Port Royal stress test at least.  If anyone is wanting a starting point for a stable OC, probably a good place to start.  I don't think my card will go much more as it crashes in PR at +165 sometimes.

Just got my 3080ti FE yesterday. Bought it from a friendly neighbor who managed to get a 3090 Kingpin, lucky. Only 500 over msrp not bad for walking 50 yards to buy it, lucky too. Anyways got it hooked up with a 10850k @ 4.8ghz on a h100 AIO(240mm), 32gb ram @3200mhz, 1080tie on afterburner +75/ +300. I know I can clock my GPU higher but it was such a long wait to get it I just want to baby it. Time Spy: 18746   GPU score: 20892
 
PS: new member today, so Ill post link in 5 days...
 

 
 
 
 
 
 

Here are my 3080ti FTW3 results, card is at stock settings, haven't done any overclocking yet, just seeing where everything stands with system as it is currently. LOD prompt due to 16x Aniso and LOD Clamp: High Qualty in nVidia control panel.
Running:
i9 9900k at 4.9ghz - 4.6ghz cache
16gb Ram at 4266mhz - timings 17, 17, 17, 36, 427 2t.
 
Edit: Guess I'm too new for links. Scores in place of links.
 
                               Time Spy Extreme                                                                          Time Spy
 
3DMark Score:           8785                                                                                             17512

Graphics Score:          9981                                                                                            19532

CPU Score:                5234                                                                                             11042

Graphics Test 1:         63.58 fps                                                                                      128.62 fps

Graphics Test 2:         58.42 fps                                                                                      110.99 fps

Avg sim time/frame:   66.9 ms                                                                      CPU Test:   37.1 fps

Hmm... it may be time to upgrade the rest of my system.  I'm able to run Time Spy at +160/+1125 which should be good for a graphics score of 20k+, but I end up with a score of 15885 because the entire rest of my computer is a bottleneck.  Hopefully Alder Lake drops in October as the rumors have been saying.  
  
https://www.3dmark.com/3dm/64186215?  
  

 

This is my best score so far http://www.3dmark.com/spy/25651150, 19,913 for graphics.  With me running a Threadripper 2950x and keeping everything except for the GPU at stock, I am definitely limited in my pursuit of a higher Time Spy score.  The overall overclock I was able to achieve is very solid and I would say I have pretty good silicone.  I won't be taking any records, except the best GPU score for my combination of HW (2950x paired with a 3080 Ti).  I have learned a lot about tweaking this card over the last week. 
 
For starters, I really don't like this boost technology and wish we could go back to just doing normal OC without explicit mods on the cards.  Setting a curve as the means to OC is unreliable since the existence of the curve is no guarantee that the card will in fact obey the curve.  Many times boost technology may overshoot the frequency for a given voltage even if the curve supposedly sets an upper limit resulting in a crash or instability.  Even limiting upper clock speeds using nvidia-smi -lgc low,high command is not much help because the overshoot can happen at any VF point. 
 
I have found that it is essential to test both light and heavy load, as well as ensure that testing is done under lowest and highest temperatures you expect the card to get to when operating.  Typically this means that I would end up setting a curve under heavy load (extended looping of Time Spy graphics test 2 for example), then save the curve, then let the system idle for a while, apply the newly set curve and see how much the system automatically ends up binning up, and then finally reduce the binning down to a number that would work by shifting the entire VF curve down and to the right (lower frequencies at higher voltages). 
 
Shifting the curve under no load ensures that the system will stay stable at both high and low loads, or at low temperatures when the built in boost algorithm may decide to overshoot the curve.  Finally, I found that simply using an offset curve (like +165 core) can produce better results in some cases than aiming to maintain a high boost consistently, while in other cases, maintaining a high and consistent boosts is better.  This is more application specific and doesn't affect the overclock so much as it affects the fact that a single OC profile may not be the best for everything. 
 
Overall, for day to day use, +165/+1,200 seems to produce good results and has so far been resilient to crashing.  For benchmarking I can comfortably do +195/+1200 with no artifacts or crashing.  I am sure some lighter games could manage with this, but I did not find such high overclock to be stable in general (I cannot indefinitely loop Time Spy graphics test 2 for instance).  Temperatures under full load stay in the upper 40C range when my room is maintained at a normal 21-22C ambient, and low 50C (52C-53C) when the ambient is around 25C.  This is cooled using an EK Vector water block.
 
The last thing I want to point out for anyone reading is this - when overclocking these cards there truly isn't a combination that is 100% stable.  Any overclock could result in an unpredictable overshoot in some scenario that one didn't test.  It is impossible to determine what that is or when that is.  A stable overclock is only generally stable, and that is all.  It may still experience issues if facing a specific set of circumstances, in a specific workload.  This is just an inherent limitation of the boost algorithm as it is today.