670 Overclocks

[Deders]

After discovering my 670's memory will quite happily overclock to 7GHz (1725 actual clock) which has improved performance in the few games I've tested so far. I'm just wondering what kind of overclocks others are getting out of their 670's.

I have seen some that seem to run 500MHz over this within the 3Dmark results page but I'm not sure if they were clocked purely for benchmarking or if they run that speed all the time for gaming.

I've tested my overclock with Sisoft Sandra's Vram latency and it seems they must increase the timings to get the cards to clock this far but the gaming results speak for themselves. Arkham City averages at 60fps with Physx and 16xQ AA. Before it was closer to 55.

For the record the max the GPU seems to want to go to is 1163. no matter what minimum voltage I throw at it, It generally seems to work at 1.175 anyway, sometimes clocking down a notch for some reason.

I've had to set the minimum voltage to 1.012 in order to get Tomb Raider to not glitch in one or 2 places.

I've set the fans so that they keep the card under 70c no matter what, and it's not too noisy. Can't hear it over general gameplay.

Generally the drivers report that the operating voltage and the reliability of that voltage are what's keeping the GPU at 1163.

Has anyone unlocked the max voltage and what would be your experience of it if you have?

 

[wand3r3r]

NVidia purposely neutered the cards and forced the card makers to comply or they lost the warranty (probably to ensure you don't OC to the next model up and instead pay the ransom). If you mean the voltage unlocking BIOS, it voids your warranty. (Not apparent if you know this)

In one card it went from around 1280 core to ~1400. This was before NV started forcing the neutering of their cards last fall...

Check your FPS etc. as you OC your memory, there are certain clocks which are worse than others and it doesn't go linearly up, in fact it can drop at certain frequencies.

Here's an excellent resource to get you up to date on the memory hole / sync / error correction issue. It's not straightforward this generation.
http://www.thetechbuyersguru.com/VRAMocing.php

(1) the video card is auto-adjusting the timings of the VRAM: this theory suggests that to maintain stability, the video card could be raising ("loosening") the VRAM timings, which in combination with the VRAM clock frequency determine overall performance.

(2) the video card core and memory clock are "out of sync": this theory suggests that memory clock operates best when "in sync" with the GPU clock, and at a certain VRAM overclock speed, it is most out of sync.

Just because your memory clocks are "higher" doesn't mean it actually performs better. Note 7Ghz above. Also, it appears to vary per game etc.

 

[Deders]

I was reading that article yesterday which is what has got me intrigued this morning. I'm currently testing with Sisoft the difference in latency and bandwidth at 100MHz increments from stock. I could upload the results when I've finished.

I do remember having a NForce 2 board which gave much better memory results when the memory controller was in sync with the CPU and I can imagine that shader clocks being in sync with the GPU clock would be of some benefit.

I might test to see how close I can clock the Memory to be in sync with the GPU but as the GPU clock is dynamic and the memory clock is static it shouldn't really affect framerates too much.

As for voltage unlocking I seem to remember there being a software that could do this for my old 560TI, but that was EVGA with an unlimited warranty, My current card is second hand so not sure whether I have any rights to a warranty or not, maybe the first user didn't register it with Asus. Still I wouldn't like to break it.

 

[toyota]

lol if you actually look at the frame rates instead of the exaggerated graphs, its fractions of 1 fps difference for Nvidia where it drops off. oh no I got 48.6 instead of 49.3. that will be within margin of error if running benchmarks anyway so not that big of a deal.

 

[Deders]

Yeah I can see that, at the moment I'm just looking for at which points the card loosens it's timings.

In the same way of thinking if boosting the memory can get me 60fps more often than not then things do become noticeably smoother.

One interesting thing I just read which makes sense: Unofficially, NVIDIA’s GPU Boost power target for GTX 670 is 141W instead of the 170W TDP. I was wondering why my computer was drawing quite a lot less power than my old 560TI (which was also rated at 170W TDP) when playing intensive games at higher framerates.

It just so happens that 141*122%=172.02, 122% being the maximum power target the Asus GPU tweak will offer me. Furmark sometimes rises to 133% potentially drawing 187.53W into the card, but I guess most of the actual chip (i.e a large portion of the shaders, tessellation units etc) isn't actually in use.

I'm wondering if the DirectCU II and the TOP edition have the same VRM's.

Has anyone flashed a non TOP edition with a TOP bios?

 

[wand3r3r]

lol if you actually look at the frame rates instead of the exaggerated graphs, its fractions of 1 fps difference for Nvidia where it drops off. oh no I got 48.6 instead of 49.3. that will be within margin of error if running benchmarks anyway so not that big of a deal.

Yeah it's not a huge decrease, but it decreased with a higher clock instead of increasing!

The point is to increase awareness about the fact overclocking the memory has unintended affects/defects.

On top of that it was 1 game shown here in the thread (a few in the link) and different games depend on memory to varying degrees so it's not saying there's a 1% decrease across the board going from 6800MHz to 7000MHz. A person has to document the progress as you overclock to find the best clocks. I'm not sure what you have against that proven concept.

With the neutered keplar cards being given hard limitations to ensure you can't overclock much, it is critical to test the core and memory separately to see the varying effect. If you are limited by TDP it could be wise to limit the memory clock to allow the core to be higher. If the higher memory clock is actually worse, it would be detrimental to try clock it higher and force the core lower. (Where limited by TDP/voltage)

I was reading that article yesterday which is what has got me intrigued this morning. I'm currently testing with Sisoft the difference in latency and bandwidth at 100MHz increments from stock. I could upload the results when I've finished.

I do remember having a NForce 2 board which gave much better memory results when the memory controller was in sync with the CPU and I can imagine that shader clocks being in sync with the GPU clock would be of some benefit.

I might test to see how close I can clock the Memory to be in sync with the GPU but as the GPU clock is dynamic and the memory clock is static it shouldn't really affect framerates too much.

As for voltage unlocking I seem to remember there being a software that could do this for my old 560TI, but that was EVGA with an unlimited warranty, My current card is second hand so not sure whether I have any rights to a warranty or not, maybe the first user didn't register it with Asus. Still I wouldn't like to break it.

Do upload the results, it will be interesting to compare them.

Asus cards are indeed tied to the serial, I'd be cautious flashing the BIOS if the only gain is the TOP clocks. MSI Afterburner should allow the same clocks with just a slider and you won't void the warranty. (You already had them as you mentioned the OC)

 

[Deders]

With the neutered keplar cards being given hard limitations to ensure you can't overclock much, it is critical to test the core and memory separately to see the varying effect. If you are limited by TDP it could be wise to limit the memory clock to allow the core to be higher. If the higher memory clock is actually worse, it would be detrimental to try clock it higher and force the core lower. (Where limited by TDP/voltage)

So far overclocking the memory by 233 actual MHz (892 effective) has made no difference in the limit to in game clock speeds.

Am I right in thinking that the memory doesn't run off the 12v rail anyway? I seem to recall that in the past it was drawn ffrom either the 3.3v or 5v lines in the smaller section of the slot.

 

[3DVagabond]

The 24pin+8pin that go to your mobo are off of the 12v rail.

 

[Deders]

The 24pin+8pin that go to your mobo are off of the 12v rail.

Yes but the PCIe slot itself provides 3.3v and 5v as well as another 75W of 12v power.

 

[3DVagabond]

Yes but the PCIe slot itself provides 3.3v and 5v as well as another 75W of 12v power.

I'm not so sure about that. The only rail that's supplying power to the mobo though, and therefor anything that the mobo is powering, is 12v.

 

[Deders]

I'm not so sure about that. The only rail that's supplying power to the mobo though, and therefor anything that the mobo is powering, is 12v.

The 20+4 pin connector supplies a variety of voltages to the motherboard which are then passed on to various components, including the PCIe slots which are specified to carry 3.3v, 5v, as well as 12v. Also if you take into account the VRM's, you'll find that lots of things on the motherboard aren't 12v, my CPU for example runs between 1.272v and 1.288v depending on the load, a tenth of 12v.

Edit: Although on testing, Furmark does seem to knock the GPU power and voltage down a notch when I increase the memory clock, but then the improvement I get with the extra 1000 effective MHz in the memory far outweighs the 13MHz I lose by implementing it. Also Furmark seems to be the only app that actually makes the driver downclock the card due to not enough overall power. I doubt this would happen in games.

MSI Afterburner should allow the same clocks with just a slider and you won't void the warranty. (You already had them as you mentioned the OC)

I've Just tried MSI's afterburner but it seems the minimum voltage is greyed out so there's not much to play with on non MSI cards.

I've got a few things to get done but I'll post the results when I get round to finishing them.

 

[3DVagabond]

The 20+4 pin connector supplies a variety of voltages to the motherboard which are then passed on to various components, including the PCIe slots which are specified to carry 3.3v, 5v, as well as 12v. Also if you take into account the VRM's, you'll find that lots of things on the motherboard aren't 12v, my CPU for example runs between 1.272v and 1.288v depending on the load, a tenth of 12v.

Edit: Although on testing, Furmark does seem to knock the GPU power and voltage down a notch when I increase the memory clock, but then the improvement I get with the extra 1000 effective MHz in the memory far outweighs the 13MHz I lose by implementing it. Also Furmark seems to be the only app that actually makes the driver downclock the card due to not enough overall power. I doubt this would happen in games.



I've Just tried MSI's afterburner but it seems the minimum voltage is greyed out so there's not much to play with on non MSI cards.

I've got a few things to get done but I'll post the results when I get round to finishing them.

Actually, I have to take back what I said. Researching it a bit, it appears that the 24 (20+4) pin plug does draw power from the other rails as well. Sorry, for the bad info.

 

[wand3r3r]

As for the tdp being hit and affecting the choice between memory clocks vs core clocks, I believe it's more apparent on the Titan / 780 which are close to the limits and don't have much room for overclocking due to the voltage neutering.

 

[nightspydk]

Yes but the PCIe slot itself provides 3.3v and 5v as well as another 75W of 12v power.

..and that is the truth. On some boards you can actually cut that power if so wish.

(runs directly thru the pci-e slot)

 

[Deders]

..and that is the truth. On some boards you can actually cut that power if so wish.

(runs directly thru the pci-e slot)

Do you mean the 12v? I thought that each connection on a gfx card had to be electrically isolated from each other in case someone used a PSU with actual separate rails, instead of one big rail being divided into several. Otherwise if the wave pattern in the electricity met at different points and didn't line up something would go bang.

If you mean the other voltages surely not every piece of hardware would have the VRM's auto tune depending on what voltages were supplied.

 

[Deders]

Ok here are the results of the latency and bandwidth tests. I used Sisoft Sandra's default tests for each frequnce going up from default in steps of 25MHz which are reported in the Asus tweak software as 100MHz steps as GDDR5 effectively quadruples the amount of data you can pass through it when compared to SDR memory.

• The Reported bandwidth was from the GPUz reading, showing it's maximum theoretical potential
• The Sisoft bandwidth results are using the DX11 benchmark, no point in using DX10
• The Sisoft Latency benchmark i used was the OpenCL one as i didn't realise there was a direct compute benchmark until halfway through, and even then results gave higher latencies so I decided to stick with OpenCL.
  • The results are an average of 16 tests ranging from 2KB to 64MB. anything from 32k and below could fit into the L1 cache and anything between that and 512 could fit in the L2 cache meaning the results were much faster than anything above that.
  • The default paged random access pattern was used.
  • You'll notice there is also a speed factor result that I included from the Latency results, again the lower the better.

You may also notice the default speed as reported by GPUz differs slightly from the rest of the frequencies as it has an additional 0.3MHz which implies a 1.2MHz effective increase.



The first thing you may notice is the increase in latency immediately above stock frequencies. The situation improves slightly up until 6808MHz where the increase in MHz cancels out the effect of the higher latencies and the speed factor is the same as it was at default.

I've not included it in the results shown here but the speed factor stays the same for 6898MHz before the next increment up to 6908MHz where there is a slight anomaly, I was sure to retest anything that didn't fit into the curve but in this situation it seems that 6908MHz is less than optimal. anything after that gave a better clock to latency ratio than the default clock.

The same occurs at 7408 before improving again., there is a hidden result (7448MHz) that I will explain in the next paragraph which also gave a speed factor result of 13.1

I was surprised to see that clocks were getting this high so I decided to test for artifacts with EVGA's OCScanner which has a test using OpenGL where it fills the memory before displaying a furry block to heat up the GPU, and sure enough at 7508MHZ it was artifacting quite badly. I had to reduce it to 7448 before I was happy there were no artifacts.

Time to start benching some games:

Tomb Raider came first and was quite happy with 7448MHz, producing the fastest results I've seen. 42FPS (2fps over 7108MHz) with no sign of artifacting.
This was with everything maxed out including TressFX and 2x SuperSampled AA (@1680x1050).

Metro Last Light on the other hand started to produce artifacts at anything over 7108MHz.
These were quite easy to spot as they were large (for artifacts) blue circles that were about a quarter of an inch in diameter.
One thing I noted as I gradually brought the clocks down was that there was no change to the performance so memory bandwidth isn't the bottleneck on my machine.

Arkham City seemed to provide the smoothest/most reliable results when the memory was clocked down again to around 6808-6898, giving higher minimum framerates and consistently ending up with 59-60 average FPS. Increasing the memory clock would sometimes give a lower average even though the max FPS were sometimes higher.

Bioshock Infinite slowed down considerably when clocked any higher than 7008MHz, indicating memory faults that had to be recalculated.

 

[Silverforce11]

Becareful with vram OC, in my experience, if you push it and it starts to artefact, it may become unstable at lower clock speeds down the road.. it sorts of degrades a bit when pushed too much. Core OC tend not to have this issue.

 

[Termie]

Excellent research, OP. But as you found...

Be careful with vram OC, in my experience, if you push it and it starts to artefact, it may become unstable at lower clock speeds down the road.. it sorts of degrades a bit when pushed too much. Core OC tend not to have this issue.

You're theoretical testing was ideal for determining the relationship between latency and bandwidth, but it wasn't effective for identifying a useful overclock. Your in-game testing led to the same conclusion I've had - 6800MHz is about the highest stable overclock that results in positive scaling in all games.

 

[Deders]

Yep, I think I'll stick to around that, seems to be the one that is consistently smoother, maybe something to do with the speed factor being the same as the default clocks, might be more in sync with the rest of the card.

 

[Praxis1452]

That was really interesting. My mem only goes +225(or 1625), +250 gets me artifacts pretty quickly.

I used EVGA precision to OC and over volt. I'm pretty comfortable flashing a new BIOS so I used a BIOS editor to set my power target to 200W and max Power to 225W which is the limit of a 6 pin + 8 pin IIRC. to be honest It never gets close to 200W even with furmark and the max voltage I can give it. Stays right around 79% TDP so ~160W. I've seen it rise up to 170W at 87C once. The higher the temp goes the greater the TDP. My fan profile keeps it below 70C though it is loud on the reference cooler.

With my max voltage locked to 1.212v (I think) I can hit 1293 as my max boost clock. My base clock is 1200 also modded in the bios.

I can actually benchmark 20-30mhz higher on the core, but I will see some graphical artifacts. Funny thing is I know that I can hit 1280mhz on the core with just 1.175v. It passes unigine heaven at that setting. Doesn't really run much hotter with my fan profile, and the noise is about the same with the different voltage settings.

By the way, when you say that your max core clock is 1163, do you use gpu-z to check max clocks during a benchmark? It can record core frequency and you can set it to max to determine the actual max clock, because if 1163 is base, it can still boost quite a bit higher into the mid 1200's. Also the downclocking is due to temperature or TDP. Almost surely temperature. Most cards jump down a boost bin at 70C, although mine is weird and doesn't do that. In fact with my current BIOS settings, it doesn't even drop a boost bin 85C.

 

[Deders]

By the way, when you say that your max core clock is 1163, do you use gpu-z to check max clocks during a benchmark? It can record core frequency and you can set it to max to determine the actual max clock, because if 1163 is base, it can still boost quite a bit higher into the mid 1200's. Also the downclocking is due to temperature or TDP. Almost surely temperature. Most cards jump down a boost bin at 70C, although mine is weird and doesn't do that. In fact with my current BIOS settings, it doesn't even drop a boost bin 85C.

I'm using the Rivatuner statistics server that comes with Precision to give me on screen real time info like core frequency, voltage, temperature, fan speed, memory frequency usage and Memory controller load amongst other things. 1163 is the maximum I see in both this and GPUz

The newer drivers have implemented a secret feature that can be read by GPUz as well as the OSD Riva offers that tell you why the card clocks down.

It can either be:

1, The total power used by the card has reached it's max (for some reason instead of the TDP Nvidia have set it for 141W, but if you add the 22% overhead that most overclocking software gives you, you get 170.2W, which just happens to be the actual TDP, obviously Nvidia want us to overclock them.)

2, Thermal limiting like you said, it will clock down 13MHz if it reaches 70c and even more if it hits 80c.

3, the operating voltage

4, the reliability of the voltage

5, or limited by GPU utilisation. This usually happens during loading screens and FMV cut scenes etc when the card doesn't need to use as much power, the driver tells it to drop down to stock or lower, I noticed that Dishonored played fine at 7xxMhz or so, I seem to remember playing through Homeworld on my old 560TI and it only rose from desktop clocks (50MHz) when I had over 100 or so ships in my fleet on screen, even then it only rose to 425MHz instead off the full 900.

My reading is usually a combination of 3 and 4 (unless I'm using something like Furmark in which case no 1 comes into play) this is typical whether I've overclocked or not.

Like you I've set my fan controller to never let the card go above 70c so temperature is never flagged. I'm tempted to ask what you use to edit and reflash the bios, I'm pretty sure that My card could take a few extra volts but the new version of Precision doesn't seem to let me unlock this anymore.

Would it affect desktop voltages etc or can you just adjust the gaming profile?

Also iirc 75W from the slot plus 2x 6 pins @75W is 225W, swap one of those for an 8 pin and you have the potential for 300W but I'm not sure if the card would need it for general usage. Like you say they are very efficient and rarely approach the actual TDP unless really pushed with something like furmark.

 

[toyota]

wow 1163 is really low for max oced boost on a non reference card like that. all i did was raise the power target on my EVGA 670 and it hit 1189 out of the box.

 

[Deders]

yeah that's what I'm thinking, I've not raised the power target too much as it seems to max out at 1.175 anyway. Might give it a go when I get home.

 

[toyota]

yeah that's what I'm thinking, I've not raised the power target too much as it seems to max out at 1.175 anyway. Might give it a go when I get home.

um raising the power target to max is the very first thing you should do. it does no harm at all to do that but not raising it will simply limit your max boost especially when overclocking.

 

[Termie]

yeah that's what I'm thinking, I've not raised the power target too much as it seems to max out at 1.175 anyway. Might give it a go when I get home.

Voltage and power are not the same thing. Raise the power limit or your voltage and your overclock will do nothing. The card is limited by power at stock settings.