GTX 670 VRAM Overclock

Joined
Sep 14, 2012
Messages
6
I might be completely off my rocker (not surprising)- but my vram appears to run at 7.6ghz with no trouble at all. Kept going up by +50 for quite some time- then figured something was a bit nutty so I set the memory to +1000 in Afterburner. Started artificing, so I turned it down to +800. No more artifacts.

(A little backstory: I wasn't happy that my DirectCU-II card had naked VRAM chips- so I made some modifications to my card. The chips on the front are now cooled by the GPU heatsink, the ones on the back are passively cooled by the backplate)

I don't know if I'm destroying my memory chips by putting such a huge overclock on, or if I'm on to something big here. Could someone please tell me if I'm just being silly (wrong clock readings) as I'm a big newbie when it comes to Kepler overclocking.
 
Just use the monitoring feature in Afteburner to check your memory clocks. My Gigabyte 670 runs at 7516Mhz stable so its certainly possible.
 
Couple of things to consider-

1. GDDR5 has an error correction function built into it, so while you might get a "crazy" overclock, if it is too high, the effective memory bandwidth drops as it corrects for errors. Therefore, your performance at 7.6ghz might be worse than it is at 6ghz. That's a trial and error type thing as you play with it - you basically want to increase it until your performance starts to decline,then back it off a little bit.

2. On GTX 670/680 series cards, you tend to get a better return on performance by overclocking the GPU at the expense of the memory. Overclocking the memory uses board power that could be used to put into the GPU, so a 7.6GHz memory overclock will likely limit your GPU overclock if you follow.

We do overclocking for these cards by finding the maximum GPU frequency first, and then boosting the memory until it starts to hit what the maximum GPU clock was set to. For 660Ti's, we take more of a balanced approach, as boosting both makes the best gameplay impact....
 
1. I've yet to see any performance decrease on either of my GTX 670's from raising memory speed. I'm aware of the supposed error correction but I haven't seen it affect performance yet on either of my samples, both continue to perform better in multiple benchmarks until the memory finally crashes. Same happened with my AMD 7850 and 7950 also, I never saw a decline in performance from raising memory speed, until it finally crashed due to memory instability.

2. While I somewhat agree with the board power that you're talking about, I've yet to see my core speed go up any higher from lowering memory speed or even leaving it at stock. Power has not been an issue on my cards, it usually that the GPU simply won't overclock any further regardless of memory speed, at which point I can raise the memory speed as high as I want. For example, my best Gigabyte 670 does 1270 max boost core speed, 7516 memory speed, lowering the memory to even the default speed of 6000MHz doesn't change my my core overclock at all, still goes no further than 1270MHz. And while I tend to agree that in years past core overclocking generally does lead to much larger improvements than memory overclocking, I don't think that the 670 and 680 are the same. I've seen very good improvements from memory overclocking on these cards and I tend to think its because of the limited memory bandwidth so they simply respond very well to memory overclocking.

Just my personal experience on the cards I've owned and I've done more benchmarks that I care to ever count.
 
1. I've yet to see any performance decrease on either of my GTX 670's from raising memory speed. I'm aware of the supposed error correction but I haven't seen it affect performance yet on either of my samples, both continue to perform better in multiple benchmarks until the memory finally crashes. Same happened with my AMD 7850 and 7950 also, I never saw a decline in performance from raising memory speed, until it finally crashed due to memory instability.

2. While I somewhat agree with the board power that you're talking about, I've yet to see my core speed go up any higher from lowering memory speed or even leaving it at stock. Power has not been an issue on my cards, it usually that the GPU simply won't overclock any further regardless of memory speed, at which point I can raise the memory speed as high as I want. For example, my best Gigabyte 670 does 1270 max boost core speed, 7516 memory speed, lowering the memory to even the default speed of 6000MHz doesn't change my my core overclock at all, still goes no further than 1270MHz. And while I tend to agree that in years past core overclocking generally does lead to much larger improvements than memory overclocking, I don't think that the 670 and 680 are the same. I've seen very good improvements from memory overclocking on these cards and I tend to think its because of the limited memory bandwidth so they simply respond very well to memory overclocking.

Just my personal experience on the cards I've owned and I've done more benchmarks that I care to ever count.

1. Not all memory will behave the same- its just that GDDR5 is different from all of the prior memory out there with its ECC capabilities, so that is something to watch for even if it may or may not rear its head in a particular cards (or sampling of cards).

2. Its not that you'd see a higher overclock or higher boost speed, its that the card will hit the power wall and scale down the GPU in resposne. Brent's recent GTX 660 non-TI overclocking review describes this pretty well, but instead of seeing the 1270 boost clock while in game, you'll see it jumping all over the place which implies you've splatted on that power wall.

I'm also curious what you are calling your benchmarks - the only thing that matters to me at the end of the day is gameplay experience, and from the testing we have performed on actual games, the 670/680 cards are not at all memory bandwidth limited. Even the 660Ti seems to be well balanced from a memory bandwidth perspective. Of course, everyone's personal experience can be different :)
 
Thanks for the replies guys.

I've found that my clock speeds remain the same regardless of memory frequency- I've seen my GPU throttle down upon hitting the 122% power limit set in Afterburner- but I attribute that to 100% utilization. Doesn't seem like memory frequency makes a huge or even noticeable dent in the total amount of power consumed. Again, could be mistaken, but I've done some Kill-a-watt equivalent measurements on DDR power usage vs. frequency vs. voltage before. Nothing significant.

Also, I've heard that thing about GDDR5 and ECC before; used to have a 6950 and read that somewhere while I was doing research. Upon further recent investigation, it seems that not all GDDR5 is ECC. Apparently some Fermi workstation cards came with ECC, but word is that enthusiast cards don't. Seems logical enough as ECC is bothersome when it comes to performance, as well as overclocking (for the very same reason you described). Could anyone confirm this?

@Blkout, have you ever noticed any degradation from running your VRAM at balls to the wall speeds? Lack of artifacts and benchmark stability doesn't lie (I was looking at the Afterburner speeds when I wrote my original post- I just couldn't believe them). I'm just worried about possibly frying my shiny new $400 card.
 
your fine ive been running mine at 7.5ghz since the day i got it and its still rockin
 
1. Not all memory will behave the same- its just that GDDR5 is different from all of the prior memory out there with its ECC capabilities, so that is something to watch for even if it may or may not rear its head in a particular cards (or sampling of cards).

2. Its not that you'd see a higher overclock or higher boost speed, its that the card will hit the power wall and scale down the GPU in resposne. Brent's recent GTX 660 non-TI overclocking review describes this pretty well, but instead of seeing the 1270 boost clock while in game, you'll see it jumping all over the place which implies you've splatted on that power wall.

I'm also curious what you are calling your benchmarks - the only thing that matters to me at the end of the day is gameplay experience, and from the testing we have performed on actual games, the 670/680 cards are not at all memory bandwidth limited. Even the 660Ti seems to be well balanced from a memory bandwidth perspective. Of course, everyone's personal experience can be different :)

To address number 2, my boost clock stays at 1270Mhz all the time in intensive games, the only time I've ever seen it throttle down is when the game isn't intensive and it doesn't need full boost clock, otherwise I've never seen the board power issue. I set the Nvidia control panel to maxiumum performance and the power setting to max in Afterburner, no issues with board power from overclocking core or memory or both at the same time. You may be correct that it affects some particular board designs but I can say for a fact that it doesn't on the Gigabyte 670's of which I own two and neither have suffered from this.

I use the Metro 2033 bench, Crysis Warhead bench, 3DMark11, 3DMark Vantage, and few others I can't think of off the top of my head, I'll check my spreadsheets when I get home from work though. Everyone of these have shown consistent gains in fps and/or score from overclocking the memory on my 670's until they actually fail to pass the benchmark.

As far saying the GTX 6xx series is balanced, its a fact that they have less bandwidth than their AMD counterparts, however through memory overclocking most of it can be made up and matched and it does show gains. I'm not saying that the GTX 6xx series is starved for bandwidth, I'm just saying there are gains there to be had which shows they will use the additional bandwidth if its there.
 
@Blkout, have you ever noticed any degradation from running your VRAM at balls to the wall speeds? Lack of artifacts and benchmark stability doesn't lie (I was looking at the Afterburner speeds when I wrote my original post- I just couldn't believe them). I'm just worried about possibly frying my shiny new $400 card.

I've never really seen any artifacts on my Gigabyte 670's from pushing the memory too far, but I have on my Sapphire 7950 when it was pushed too far, I would get blue squares(artifacts) in The Witcher 2 when it close to the limit. I've also seen some flashing textures in the Heaven and 3DMARK Vantage benchmarks on the Sapphire 7950 card when pushing the memory too far, but I've never seen this on my 670's. In fact, I've never seen any artifacting on my 670's when pushing the memory too far, just the eventual crash or application failure.

For what its worth, from my own experience, you will never damage your card from overclocking it unless you're increasing voltage which only increases the chance that you could cause some damage but its not a certainty. The worst that usually happens is that the PC locks up, crashes the application, then you simply lower the clocks until it becomes stable again.
 
I use the Metro 2033 bench, Crysis Warhead bench, 3DMark11, 3DMark Vantage, and few others I can't think of off the top of my head, I'll check my spreadsheets when I get home from work though. Everyone of these have shown consistent gains in fps and/or score from overclocking the memory on my 670's until they actually fail to pass the benchmark.

As far saying the GTX 6xx series is balanced, its a fact that they have less bandwidth than their AMD counterparts, however through memory overclocking most of it can be made up and matched and it does show gains. I'm not saying that the GTX 6xx series is starved for bandwidth, I'm just saying there are gains there to be had which shows they will use the additional bandwidth if its there.

Sounds like you got a pretty good card - all of the Keplers that have been across my review bench usually had to sacrafice memory clock to acheive a higher GPU clock (including the Gigabyte 670 that I reviewed).

A word of caution with the benchmarks, especially with current generation games, is that the way they use GPUs does nto reflect actual card usage when you are playing a game. Nvidia's reviewer guides and white papers even go as far as stating a risk of damaging the card by running some of the benchmark suites when overclocked due to how they unrealisitcally use power. So while they might be able to tell you that your OC is potentially stable and effective, the only way that I would confirm it would be with several hours of gaming across a wide suite of games (which it seems like you have done).
 
Sounds like you got a pretty good card - all of the Keplers that have been across my review bench usually had to sacrafice memory clock to acheive a higher GPU clock (including the Gigabyte 670 that I reviewed).

A word of caution with the benchmarks, especially with current generation games, is that the way they use GPUs does nto reflect actual card usage when you are playing a game. Nvidia's reviewer guides and white papers even go as far as stating a risk of damaging the card by running some of the benchmark suites when overclocked due to how they unrealisitcally use power. So while they might be able to tell you that your OC is potentially stable and effective, the only way that I would confirm it would be with several hours of gaming across a wide suite of games (which it seems like you have done).



One of my Gigabyte 670's does 1270/7516MHz memory stable, the other Gigabyte 670 only goes to 1228/7200MHz stable, but I'm happy with both. Its just luck of the draw I guess, both run the same temps about 60c under full load, one just overclocks a little better.

I don't use any benches like Furmark or other burn-in type stress tests. Looping Heaven is about as stressful as it gets for 30-60 minutes at max settings without actually playing a game. The other benches are just used to get consistent results to see if its improving from the the overclocks. But in the end, my two PC's aren't bench queens so I do play a lot of games and if my overclocks aren't stable gaming either, then its no good for me.

All of these Kepler cards are all over the place in regards to overclocking, I've seen some hit mid 1300's core speed with no voltage adjustment, but their memory overclock was lacking and was only able to achieve about 6400-6700 so I guess every card is unique. Ideally I would love to have one that does about 1300 core and anything over 7200 memory but I don't.
 
What brand vram heat sinks did you use?

Also, how did you adhere them to the dcu2 670's chips?
 
On the front: there are two chips which can be "hooked up" to the main heatsink using some thermal tape. There's a 3-4mm gap between chip just above the PCI-e connector- the one opposite it has a smaller gap (around 2-3mm). I placed two Modmytoys mini-heatsinks on the edge of both of the right-hand chips. I think it would have been more effective to use the Zalman ZM-RHS1 heatsinks- those are shorter and you can probably cover more of the chip.

On the back: Took the backplate off- found that there's insulating material all over the parts that could possibly ever make contact with the vram chips. Used a bit of thermal tape in a '[' shape which connects the chips to the heatsink. I didn't remove the insulating material, though it would be more effective to partially remove the material, at least the areas right above the chips.


Ideally, you'd want to machine a bracket for the front vram chips and use a thermal pad to "connect" them to the main heatsink. It's not an easy task though. Going to get a full-cover waterblock then I can wave bye-bye to the silly DirectCU cooler.
 
That cooler was the only reason I bought the card. It is great and water cooling is too expensive to justify imo. Thanks for describing your process :)
 
Back
Top