pls help, problem overclocking a GPU

[raclimj]

my brief specs:

Intel Xeon E3-1230 V3
Gigabyte Z87X-OC
MSI GTX 780 Twin Frozr OC 3GB w/ NZXT G10 Bracket + Corsair H110 280mm AIO Liquid Cooler
Seasonic X660 Platinum

more info of the GPU:
It is revision B1
It is using Hynix Memory
I am using the latest driver (344.11)
Default Clock: 902 Mhz Core, 1502 Mhz Memory, 954 Mhz Boost
Stable Overclock w/ stock voltage(highest I saw was 1.1620): 1030 Mhz Core, 1845 Mhz Memory, 1082 Mhz Boost (1137 Mhz Max Boost while gaming)
the absolute highest gpu temp I saw while stress testing was 54c.


I tried to verify my gpu core overclock stability through EVGA OC Scanner X(artifact scanner), Furmark, Unigine Valley and I never got any artifact, flickering, etc... but as soon as I play a real game (Grid Autosport) it crashes(It never crash while on stock).


Right now the only one that seem to actually test the stability of my GPU core overclock is OCCT with error check. I was able to set the base core clock to 1030Mhz and it seems to be stable as It never crashes on game or show any weird stuff.

In comparison to OCCT other programs such as EVGA OC Scanner, Furmark, etc make it seem like my overclock is stable (1115Mhz on core) as it never shows any error, artifact, flickering, etc... on those stress test but the overclock fails on actual games.

The problem with OCCT is that it doesn't detect errors for memory stability. As an example If I overclock the memory (like 1900Mhz), It would just pass the OCCT error check (even with the option for memery used to the max) but when I play game I get a microsecond black screen so I have no reliable and efficient way of actually testing stability of memory overclock (I got my 1845 Mhz overclock by playing an actual game and reduce the memory frequency by -5 till I don't till I don't get any microsecond black screen). OCCT also doesn't load the Max Boost clock while stress testing (the max for GPU core was 1058 Mhz and 1.0870v).

I also encounter a weird problem that if I overvolt the GPU (even with the absolutely lowest added value), the performance got worse (on OCCT the FPS drops by about 15-20FPS). Adding more and more volt makes the performance drop steadily with every volt increase. The only thing I think what is causing the problem is the Power Limit/Target, My GPU only allows a maximum of 103% and I suspect I am reaching the power limit by adding voltage (changing the priority to temp does nothing). I am considering flashing another BIOS or more specifically the MSI GTX 780 Lightning BIOS as it allows up to 109% power limit but I am not sure if it is safely compatible (I heard that on some BIOS like on the R9 290 with Elpida Memory, there was a bug where if you overclock the memory, the memory timings gets loose which affects the mining hashrates performance more info: HERE) (to clarify, I am worried that different BIOS might have different memory timing setting). I also considered the Kepler Bios tool but I have no Idea how I could edit it to increase the power limit from 103% to 109% or something.


To recap I need:
-A reliable(one that actually WORK) and efficient(doesn't take 100 steps and hrs) way to test stability for the gpu core AND memory(especially) overclock
-increase the Power Limit/Target from 103% to a higher figure (like 106%, 109%) to prevent throttling with overvolting
-view GPU memory timings

 

[Araxie]

First.. try without that excessive amount of vRAM clock... thats the main problem here at 1845mhz(7380mhz) you are over 1000mhz over the stock 6000mhz vRAM clock.. Also with that card, download the MSI gaming app and change the Card Profile to Overclock mode, as some card by default are selected to Gaming mode.. but your main problem its that excessive overclock in the memory..

 

[dragonstongue]

umm to my knowledge you cannot look at gpu memory timings, you don't really need to anyways, the only way to reverse engineer this if you will find out the model number and maker of the Vram and check online you might luck out on this. Speed(clock rate) and timings are not the same thing.

I could be wrong but pushing memory that fast might be asking to much.(6008 combined stock to 7380 is a huge jump given these are more or less already top of the bin IC) Least from what I have had in my own experiences, if you want the high core start with the core, then add the memory clock after the fact, there is only so much power that can go around, and cranking the memory clock up will limit the power available to said core.

Core should able to go higher then that for sure, but in this alone you are asking for 100% stable settings that someone else uses, EVERY gpu is different part of the reason why many makers do not massively clock the card from factory, small bump 100% will work, but what you are doing is not a small bump.

And while yes it sucks to have to test and test and test, that's part of the game takes time to do this as overclocking is not a guarantee. I have had twins of a few cards before and one will always clock higher and push lower heat then the other, other will need more volts to hit same temps etc. So you can rush to find the answer or get the "best" answer by doing what you are, test, test, test.

Or well don't overclock it.
ditto on araxie LOL

 

[runudownquick]

^^We posted at the same time & I totally agree^^

I agree about the memory. You've taken it from a stock 6Ghz to 7380Mhz...that's a huge and impressive leap with the hynix chips.

Take off the memory overclock and try everything again. You should be a lot more stable and have a lot more headroom to clock up your core. Once you find the limit of the core, slowly push up the memory as well. If you're just running a single 1080/1440 monitor than a sky high memory overclock will be less important than in running 4k/surround monitors.

 

[raclimj]

First.. try without that excessive amount of vRAM clock... thats the main problem here at 1845mhz(7380mhz) you are over 1000mhz over the stock 6000mhz vRAM clock.. Also with that card, download the MSI gaming app and change the Card Profile to Overclock mode, as some card by default are selected to Gaming mode.. but your main problem its that excessive overclock in the memory..

umm to my knowledge you cannot look at gpu memory timings, you don't really need to anyways, the only way to reverse engineer this if you will find out the model number and maker of the Vram and check online you might luck out on this. Speed(clock rate) and timings are not the same thing.

I could be wrong but pushing memory that fast might be asking to much.(6008 combined stock to 7380 is a huge jump given these are more or less already top of the bin IC) Least from what I have had in my own experiences, if you want the high core start with the core, then add the memory clock after the fact, there is only so much power that can go around, and cranking the memory clock up will limit the power available to said core.

Core should able to go higher then that for sure, but in this alone you are asking for 100% stable settings that someone else uses, EVERY gpu is different part of the reason why many makers do not massively clock the card from factory, small bump 100% will work, but what you are doing is not a small bump.

And while yes it sucks to have to test and test and test, that's part of the game takes time to do this as overclocking is not a guarantee. I have had twins of a few cards before and one will always clock higher and push lower heat then the other, other will need more volts to hit same temps etc. So you can rush to find the answer or get the "best" answer by doing what you are, test, test, test.

Or well don't overclock it.

First of all, the overclock on the GPU core is the absolute max stable on stock voltage(anything higher than that, It would give 1-2 errors every 10sec which is only detectable through OCCT error checking (as I said on OP, other stress test software wouldn't even detect any errors or artifact way past my current overclock and make it seem a stable oc but it crashes on actual games which leads me to believe OCCT actually WORKS as a stability test).

As for the memory overclock, It is the the last thing I tinkered with and it doesn't affect the GPU core clock stability as I found the absolute max stable core clock BEFORE i start overclocking memory. As for the memory overclock stability, I cannot say certainty that it is 100% stable as the only method I found to stress test it is to fire up a game and see for microsecond black screen, if so I reduce the memory frequency by -5Mhz till it is no longer doing any weird stuff like black screen.

As for the original purpose of this thread, it is still open as I still haven't found a way to reliably and efficiently stability test memory overclock and increase the Power Limit/Target Past 103% to prevent throttling with overvolting.

 

[Araxie]

Power limit in your card should be 110.. not 103.. try switching to Overclock mode with the MSI gaming app to see if are capping you to reach higher power target... also the best way to test a GPU overclock stability always will be Just gaming...

 

[dragonstongue]

I know for myself that happened to be Diablo 3, World of Tanks, BC2/BF3 and Skyrim. Diablo seemed very sensitive to memory speed, where BC2/BF3 was very sensitive to core speed/volts.

If it is overclock mode or maybe you got a "bum" card that the core doesn't want to clock up though the memory sure seems to want to XD

 

[raclimj]

Power limit in your card should be 110.. not 103.. try switching to Overclock mode with the MSI gaming app to see if are capping you to reach higher power target... also the best way to test a GPU overclock stability always will be Just gaming...

MSI gaming app has nothing to do with power limit, its just a preset profile with clock speeds.

The power limit on my hard is coded into BIOS.

Spoiler: click to view

GPU Device Id: 0x10DE 0x1004
Version: 80.80.21.00.1C
GK110B P2083 SKU 21 VGA BIOS
MSINV298MS.150
Copyright (C) 1996-2013 NVIDIA Corp.
GK110B Board - 20830021
Board power target: 280.0 W
Board power limit: 291.0 W, -46%, +4%
Boost clocks
P0 Boost: 954.00 MHz
P0 Base: 901.50 MHz
Performance Level 0
Core Clk: 324.00 MHz
Mem Clk: 162.00 MHz
Performance Level 1
Core Clk: 540.00 MHz
Mem Clk: 405.00 MHz
Performance Level 2
Core Clk: 575.00 MHz
Mem Clk: 1502.00 MHz
Performance Level 3
Core Clk: 575.00 MHz
Mem Clk: 1502.00 MHz

http://www.techpowerup.com/vgabios/147726/msi-gtx780-3072-130910.html

also stress testing via an actual game is not the best as it is neither reliable nor efficient.

As an example, when I was overclocking my GPU core and testing stability via games, It would play perfectly fine on games such as Dota 2, Crysis 3, Skyrim but then It would crash on games such as Grid Autosport, Titanfall, etc... making it not a very reliable way as some games are less demanding than others or they use the GPU in a different way.

It is also neither efficient as by testing via games, you wouldn't know if it is really stable as sometimes it would just instantly crash, sometimes it would take more than a few hours or sometimes it would work ok for a few days then bam a crash. The time it occurs is totally random.

 

[dragonstongue]

http://www.msi.com/product/vga/N780_TF_3GD5OC.html#hero-overview
954MHz Core (Boost Clock:1006MHz) (OC mode)
902MHz Core (Boost Clock:954MHz) (Gaming mode)
863MHz Core (Boost Clock:902MHz) (Silent mode)
3072 GDDR5 6008MHz Memory
http://www.guru3d.com/articles_pages/msi_geforce_gtx_780_gaming_review,28.html

Years ago Asus tried using this in a reverse phycology way with me that overvolt and stuff like that is never a guarantee from the card maker(AMD nor Nvidia) no matter the control panel etc, as they stated like you are saying exactly it might be a feature in control panels but the BIOS is hard coded so it may not apply correctly even when set higher (soft set unless BIOS is flashed otherwise)

 

[doubletake]

Yeah most of these synthetic tests are garbage for detecting excessive memory OCs. Try running the games dragonstongue suggested, also Just Cause 2's benchmark, Left 4 Dead (1/2) if you have them, those games always break out in easily detectable artifacting for me when a memory oc is even just a few mhz too high. On the other hand, I found the Metro Last Light bench to be good for testing out core oc, will crash/lock up quickly if it's too high.