Hardware & Technical Graphics OverClocking and Benchmarking - help required.

[GunnerBill]

I just don't think I'm getting the best out of my system.

I'm looking at Overclocking the graphics card:

https://www.overclockersclub.com/guides/nvidia_gtx_1080ti_oc_guide/

The guide seems pretty easy to follow (I used to do this stuff all the time!) but it refers to benchmarking software (to track your changes). I've lost track of what's good in that world, having last used 3D Mark about a 1000 years ago! I'm not really wanting to "pay" for benchmarking software, anyone got any tips on a good free programme?

What about O/C the CPU??


All help appreciated!

Processor (CPU)	Intel® Core™i7 Quad Core Processor i7-7700k (4.2GHz) 8MB Cache
Motherboard	Gigabyte AORUS Z270X-Gaming 7 : ATX, LG1151, USB 3.1, SATA 6GBs - RGB Ready
Memory (RAM)	32GB Corsair VENGEANCE DDR4 3000MHz (2 x 16GB)
Graphics Card	11GB NVIDIA GEFORCE GTX 1080 Ti - HDMI, 3x DP GeForce - GTX VR Ready!

 

[Caramel Clown]

I use Furmark to stress test my GPU. Unfortunately my 1080Ti from EVGA isn't a silicon lottery winner. I was troubleshooting driver instability for the past 3 months or so, and it almost driven me insane. It's so inconsistent. Finally found a reason of driver crashes - unstable factory overclocks. After about a week of incremental clock changes 1MGhz at a time, I've come to -15 on the core from the factory. Which is still okay, since it's overclocked from the box. However, not exactly desirable.

 

[[VR] optimal_909]

I am usually launching User Benchmark, or the more respectable 3DMark free demo. There was a thread about 3D Mark score for reference.

As for overclocking, it depends what you are running. Elite will be bottlenecked by your GPU, so your 7700k will be just fine running standard, especially as it is prone to run hot.

For reference, I'm running:
MSI Armor 1080ti - undervolted as it has bad cooling, running at 1.95Ghz at 1.05V (IIRC) stable at 65-67C max
7700k - currently at 4.6 Ghz using the built in OC profile of my mobo

7700k overclocking is a tricky thing. It is said that it has very bad thermal paste below the lid which makes proper cooling very difficult.
I used to run it at 4.8Ghz at 1.23V, but that became unstable after upgrading to 1080ti from a 1060 6gb, so I reset it to the standard OC values (constantly switches between 4.6 and 4.2Ghz depending on load). However, CPU intensive stuff like Civ6 make it run quite hot, 80-85 degrees with spikes.

 

[[VR] optimal_909]

I use Furmark to stress test my GPU. Unfortunately my 1080Ti from EVGA isn't a silicon lottery winner. I was troubleshooting driver instability for the past 3 months or so, and it almost driven me insane. It's so inconsistent. Finally found a reason of driver crashes - unstable factory overclocks. After about a week of incremental clock changes 1MGhz at a time, I've come to -15 on the core from the factory. Which is still okay, since it's overclocked from the box. However, not exactly desirable.

I seem to have a very good PCB with biowaste factory cooling as mentioned above. Overall very lucky with the card especially as I have bought it from a bitcoin miner.
What is your clock speed?

 

[Morbad]

The guide seems pretty easy to follow (I used to do this stuff all the time!) but it refers to benchmarking software (to track your changes). I've lost track of what's good in that world, having last used 3D Mark about a 1000 years ago! I'm not really wanting to "pay" for benchmarking software, anyone got any tips on a good free programme?

If you are primarily looking to track performance changes and not test for stability, almost anything GPU limited that can give fairly consistent run to run results will do. 3DMark is still a thing and has free versions. Many games also have their own built in benchmarks or have repeatable seconds where one can measure performance.

Actual stability testing is quite a bit more involved, especially given that many dynamic clocking and error correcting features can obfuscate instabilities, which may manifest as subtle performance losses rather than crashes or visual artifacting. MemtestG80 (about six instances using 1.5GiB each, for 1000 loops) is a good way to find memory issues on NVIDIA cards, and Unigine Superposition and the FinalFantasy XIV benchmarks are both good stability tests that can be used to measure performance changes as well. 3DMark TimeSpy is pretty solid too, but cannot be looped in the free version.

Personally, while I'd also recommend using MSI Afterburner to OC a 1080 Ti, I'd suggest using a manual frequency/voltage curve, rather than the flat offsets. The 1080 TI's will generally become power limited fairly early on, so carefully limiting what voltages are applied is critical to getting the most out of the cards. Using such curves is quite a bit more involved than the offsets however, so the method in your guide may be a better starting point. You can probably get decent gains just by maxing out the power limit slider and setting up a fan curve that brings temps down...actual clock speed applied is heavily dependent on operating temp and power budget.

What about O/C the CPU??

A bit more involved with more serious risk of data corruption if you aren't careful, but the general process is not all that different (adjust and test), though power limitations will be a far lesser issue than how lucky your sample is and how well you can cool it. Do initial testing from bootable environment, so if it's especially flaky you won't corrupt your OS install.

I use Furmark to stress test my GPU.

FurMark is so power hungry on modern hardware that it's of limited use for finding game stable clocks and is more useful for evaluating power delivery and cooling.

On my 1080 Ti, I'm power throttled long before FurMark tells me anything else of use. It won't crash or artifact, even at blatantly unstable settings, because it throttles so heavy from heat or power limitations that clock speeds never reach the point FurMark can reveal any other issues.

After about a week of incremental clock changes 1MGhz at a time, I've come to -15 on the core from the factory. Which is still okay, since it's overclocked from the box. However, not exactly desirable.

If it's not stable in the state it comes in, I'd argue it's defective and should be returned.

 

[Zak Gordon]

I just don't think I'm getting the best out of my system.

How so? Is it just a gut feeling or are you seeing bad performance? You should be fine really. At some level if you are experiencing performance issues it will likely come down to software. The OS (esp Windows 10) can be an issue these days, and drivers for GPU's have been hit and miss for many years now.

To put it another way when i look at your hardware i can not see any reasonable weak point. So any problems you might be seeing are likely software issues (badly optimised code, conflicts etc).

Keep in mind that even the highest end hardware is not as good (or better!) than real life, we are still far way from that technologically speaking.

But run some bench tests (just watch the temps on your 1080Ti - they run hot). The Guru3D site might have some pointers here on what you could use to get some benchmarks:

https://forums.guru3d.com/forums/benchmark-mayhem.32/

 

[GunnerBill]

How so? Is it just a gut feeling or are you seeing bad performance? You should be fine really. At some level if you are experiencing performance issues it will likely come down to software. The OS (esp Windows 10) can be an issue these days, and drivers for GPU's have been hit and miss for many years now.

To put it another way when i look at your hardware i can not see any reasonable weak point. So any problems you might be seeing are likely software issues (badly optimised code, conflicts etc).

Keep in mind that even the highest end hardware is not as good (or better!) than real life, we are still far way from that technologically speaking.

But run some bench tests (just watch the temps on your 1080Ti - they run hot). The Guru3D site might have some pointers here on what you could use to get some benchmarks:

https://forums.guru3d.com/forums/benchmark-mayhem.32/

I am using Rift in VR.

Specifically my settings for planetary textures are pretty low and nudging them anywhere even slightly off the floor gives loading and framerate issues. HMD res is at 1.0 x and again increasing gives issues.

However my thread was prompted by this thread which had resurfaced: https://forums.frontier.co.uk/showthread.php/454153-Graphics-settings-Beyond-Ultra

Where some users seem to be running the game at all sorts of crazy setting, with similar setups to mine very smoothly...


But to be fair generally I am very happy with the performance, but now I MUST HAVE MOAR!!!!!!11111101010!!)!!)011111


So here I am!

-------------

@Zak Gordon.

I was going with the OCC article on Benchmarking software until I saw that Unigine was 2009. However this still seems to be a thing on the link you posted to Guru of 3D. Is this still good software?

 

[Caramel Clown]

I seem to have a very good PCB with biowaste factory cooling as mentioned above. Overall very lucky with the card especially as I have bought it from a bitcoin miner.
What is your clock speed?

It clocks to 1940 at full load, if I remember correctly. That's after my -15MGHz on the core fix. So, it's still respectable. At 106% power delivery, no changes to voltage. It reaches it's temp equilibrium at about 68 degrees, with moderate fan speeds. So, it's not all doom and gloom.

If it's not stable in the state it comes in, I'd argue it's defective and should be returned.

Well, yes and no. Since GPU is provided by NVidia they can only guarantee that it is stable at stock speeds. While manufacturers generally support overclocking, there's no guarantee it will provide tangible results. Since EVGA ships this card with pretty generous overclock I'd argue they are not breaking any rules. Probably not sufficient testing of their overclocks, I'll agree with that. Card isn't defective, it's just not lucky to be a finest example of silicon lottery There's no artifacting, power problems, BSODs. Just driver crashes. Standard sign of unstable overclocks. In fact, I was able to push it to 2040 or so, which is quite high for 1080ti. It can go like that for days with no issues, then start destabilizing.

 

[Morbad]

Well, yes and no. Since GPU is provided by NVidia they can only guarantee that it is stable at stock speeds. While manufacturers generally support overclocking, there's no guarantee it will provide tangible results. Since EVGA ships this card with pretty generous overclock I'd argue they are not breaking any rules. Probably not sufficient testing of their overclocks, I'll agree with that. Card isn't defective, it's just not lucky to be a finest example of silicon lottery There's no artifacting, power problems, BSODs. Just driver crashes. Standard sign of unstable overclocks. In fact, I was able to push it to 2040 or so, which is quite high for 1080ti. It can go like that for days with no issues, then start destabilizing.

You bought an EVGA branded product that was no doubt advertised to be stable in it's out-of-box config, irrespective of it's baked in OC. If it's not, it's not living up to it's specifications, that they undoubtedly place a price premium on. If it's not stable and you didn't break it, it's a clear cut QA problem on EVGA's end.

If it needs to be modified to be stable in a system that isn't itself defective or deficient in some way, they are breaking all sorts of rules. The very fact that underclocking it resolves the issues shows that the factory OC should not have passed testing. It can be a relatively good 1080 Ti sample and still be defective if it's marketed as something even better.

 

[Caramel Clown]

You bought an EVGA branded product that was no doubt advertised to be stable in it's out-of-box config, irrespective of it's baked in OC. If it's not, it's not living up to it's specifications, that they undoubtedly place a price premium on. If it's not stable and you didn't break it, it's a clear cut QA problem on EVGA's end.

If it needs to be modified to be stable in a system that isn't itself defective or deficient in some way, they are breaking all sorts of rules. The very fact that underclocking it resolves the issues shows that the factory OC should not have passed testing. It can be a relatively good 1080 Ti sample and still be defective if it's marketed as something even better.

Knowing how EVGA's warranty and RMA process works I could get it replaced in no time. Not sure I want to though. Factory overclocks not being stable are quite widespread, and some manufacturers probably won't even replace such cards. Pretty much every company out there will ship refurbs 95% of the time for your returned card anyway, so you can get something worse. Not to mention that 1080Ti is out of production, so it's an interesting question what I could get it replaced with.
Troubleshooting this card was madness, but my scientific curiosity kicked in and it became fun. I'm still in a process of working on it. There is a faint chance that it could be motherboard playing up, since it was the only component that wasn't replaced by something else during testing. I would much rather RMA the board if I find it being defective, since I love the card It's quiet under load, and nicely cooled.

 

[GunnerBill]

Hijack much!

lol

 

[[VR] optimal_909]

It clocks to 1940 at full load, if I remember correctly. That's after my -15MGHz on the core fix. So, it's still respectable. At 106% power delivery, no changes to voltage. It reaches it's temp equilibrium at about 68 degrees, with moderate fan speeds. So, it's not all doom and gloom.

Then it performs in a similar way as mine at higher voltage and lower fan speed - not if I needed a proof for the weak cooling my card has.

 

[[VR] optimal_909]

Hijack much!

lol

You've got Morbad's feedback, he is pretty much the authority on these pages.

 

[Jon flint]

Unigine Valley is a good one (and free), I find it quite useful to see when you've pushed too hard as you'll easily notice any artifacts/flickering on the sky:
https://benchmark.unigine.com/valley

Unigine Superposition is newer, and also free:
https://benchmark.unigine.com/superposition

3dmark is still the standard, mostly for its ability to compare your scores to other people's (with leaderboards and stats etc....), but as you said, not free for the full version. Although the steam "demo" version should give you access to the basic 3dmark timespy test:
https://store.steampowered.com/app/223850/3DMark/

CPU overclocking is a bit more tricky as you could potentially break stuff (GPU overclocking is pretty much idiot proof), but if you take the plunge, something like cinebench is a good quick test to track any increase in cpu performance:
https://www.maxon.net/en/products/cinebench/

 

[GunnerBill]

So far I've managed this as a stable bench test running the Unigine Heaven benchmark app.
Stable in game too using VR Ultra settings from DrKai.

Thanks for everyone's help, now for some game time!

----

Temp's in game were in the low 70's without excessive fan.

 

[Morbad]

So far I've managed this as a stable bench test running the Unigine Heaven benchmark app.
Stable in game too using VR Ultra settings from DrKai.

Thanks for everyone's help, now for some game time!

Looks good so far, but this illustrates where the curves could come in handy.

If you have fairly decent sample, that clock speed should be possible with about 50mV less, which in turn would reduce temps and enable higher clocks. However, you can't add a negative voltage offset, so you need the frequency/voltage curves to do this.

An example I posted a while back: https://forums.frontier.co.uk/showt...i-comparison?p=6924334&viewfull=1#post6924334

Because of my ability to precisely cap voltage for a given frequency, the curve gives me much more control over power and heat output at a given clock speed.

 

[GunnerBill]

Thanks Morbad, I got the impression from the overclockerclub article I read that although it was not possible to overvolt the card it was possible to undervolt it.

Undervolting:
Yes, you are reading this right; undervolting the card is a real thing. Why would someone do this, you ask? Well, some people like to play around with new gadgets and found that a fair share can actually run at a lower voltage and achieve the same clock speeds. On a rare occasion, over 2000MHz at 1.00v. I suspect those cards are also really good overclockers and are the gold standard of hitting the silicon lottery. It just so happens that these cards can achieve the same clock speeds of other cards that need 1.063v to do the same thing. The only reason to do this if you wanted less power draw and heat. The downside is the card isn't running at its maximum potential.

 

[Morbad]

Both overvolting and undervolting are possible, it's just that the max voltage has a fairly low cap (which is still high enough to be meaningless on an air cooled part).

To undervolt the part, you need a curve, as the offset only has positive values.

 

[GunnerBill]

Both overvolting and undervolting are possible, it's just that the max voltage has a fairly low cap (which is still high enough to be meaningless on an air cooled part).

To undervolt the part, you need a curve, as the offset only has positive values.

So it does! I guess it (the guide) means using the curves without specifically mentioning them. I think he did mention it in the video though.

It looks like I got a reasonable card, I had fun setting up the max clock and mem offsets individually; getting over 6k stable for memory on its own! I might be able to squeeze the core higher as ED is definitely not as demanding as the benchmark app was.

It doesn't seem to run too hot either, not that my experience of this is large...

 

[Morbad]

Bench carefully after OCing the memory and test it separate from the core. The first signs of instability on these (and many modern) parts is performance that stops increase as memory clock rises, and then performance actually starting to fall. Takes protracted synthetic tests (the aforementioned MemtestG80) to actually push errors through that can be quantified.