Overclock.net banner

1 - 20 of 22 Posts

·
Nvidia "shill"
Joined
·
1,429 Posts
Discussion Starter #1 (Edited)
So here's me dilemma, I have a GTX 660 in an older build I got laying around, now what i'm used to is having a large memory bus on my GPU like my 290X for example it has a 512 bit bus, so a memory overclock on that doesn't make much of a difference (actually none or maybe 2 fps across all the games I play). Now on this GTX 660 I have recently acquired since it has only a 192 bit bus should I go as far as I can on the memory overclock and ignore the core, just go straight core overclock or balance both? Now I know this may be an obvious question but let me list a few things:

1. I want to go as far as I can without bumping the voltage up a whole lot, right now sitting at a +12mv offset I would go higher but reference cooler so.... yeah.

2. In the games I have played the memory overclock has seemed to benefit more from an overclock than the core has like as in it feels like the fps is more consistent.

3. I know benchmarks don't really show real world use, but when I increased my core overclock along with the memory overclock the scores actually went down vs when it was just vram overclocked.

I guess what I'm asking is exactly how bandwidth starved is the 660 and should I just go as high as I can on the memory? From what I have noticed in the games I play I am mostly CPU bound anyway so a core overclock probably won't do much, but this is why I am asking. For those wondering the CPU is a Intel Xeon X3230 (Core 2 Quad Q6700 equivalent) it is not overclocked because I have a G series motherboard so it lacks CPU core voltage, and I am using a ghetto modded LGA 1150 stock cooler that came with my i7. I do plan on upgrading to a Q9650 in the future but for now this X3230 will have to suffice. No I will not be upgrading the platform as I have a use for this particular build and already have another, more modern platform with a Broadwell i7. Only reason I stated that as some people will recommend on upgrading the platform when in fact I actually built this around trying to see what LGA 775 is still capable of and I had spare 775 parts laying around, if you know what I mean. To be honest this is my first (ish) budget GPU so I don't really have experience on lower bandwidth cards or Nvidia to be exact, last Nvidia card I had was a Geforce MX 2 or something (whatever was in this old Socket 423 dell from 2001 that I have in my closet collecting dust.) So I mostly have experience on stuff like HD 7970, HD 4870, R9 290X, R9 380 etc (higher end stuff) Anyway any input is more than welcome and appriciated. And no, the CPU doesn't bottleneck as bad as one would think, most of the time the GPU is 80%+ in use and the CPU is rarely pegged to 100% (hovers mostly in the 60's to 80's) with CPU hammering games excluded (yes looking at you Fallout 4 and GTA V.)
 

·
9 Cans of Ravioli
Joined
·
19,918 Posts
A memory OC won't bump it up a level (GTX 670), even with a core OC. Don't waste your time, run it straight and call it a day.
maybe not, but it's still worth to OC :)
 

·
9 Cans of Ravioli
Joined
·
19,918 Posts

·
Nvidia "shill"
Joined
·
1,429 Posts
Discussion Starter #6
Yeah been messing around with it and it seems like (or at least with the games I play) the memory overclocks a lot better than the core does and just makes the fps in games feel smoother rather than being all over the place. Could be placebo effect though.
 

·
Registered
Joined
·
5,374 Posts
What Pook said on both posts^

Twilight, this is exactly why benching is so important when overclocking. On Pascal you can take the RAM too far and actually loose performance.
My suggestion:
Run heaven or valley then bump the memory as far as it'll go then write it down, reset to default and do the same to the core. Apply the core OC and run the games you play most using an unlimited frame rate while using fraps to bench it. Try to take the same path and do the same things for each run then reset to default and try with the memory.
Which ever gives you the best performance should be your priority.

If you are playing games made in the last 5 years or so I can see why your RAM might be the bottle neck as there are so many large textures that need to be stored and rendered.

EDIT: Considering the age of the GPU it might be a good idea to re-paste the card and give it a good cleaning if needed.
 

·
Nvidia "shill"
Joined
·
1,429 Posts
Discussion Starter #8 (Edited)
What Pook said on both posts^

Twilight, this is exactly why benching is so important when overclocking. On Pascal you can take the RAM too far and actually loose performance.
My suggestion:
Run heaven or valley then bump the memory as far as it'll go then write it down, reset to default and do the same to the core. Apply the core OC and run the games you play most using an unlimited frame rate while using fraps to bench it. Try to take the same path and do the same things for each run then reset to default and try with the memory.
Which ever gives you the best performance should be your priority.

If you are playing games made in the last 5 years or so I can see why your RAM might be the bottle neck as there are so many large textures that need to be stored and rendered.

EDIT: Considering the age of the GPU it might be a good idea to re-paste the card and give it a good cleaning if needed.
Yeah already ahead of you, ran both Heaven and Valley and tested Core overclocked alone, Memory clock alone and both combined and got a slightly higher score with just the memory overclock. When I received the card I spent a good hour and a half cleaning all the dust off of the shroud, pcb, vrms etc with 91% alcohol (looks and runs like a brand new card now). Repasted with AS5 tested to see If it worked and it did.

In Heaven results were:
for all tests below I used a voltage of +25mv and there was no stability issues.

Stock: 1121 pts
Core oc +65mhz: 1171 pts
Memory oc +296mhz: 1227 pts
Combined oc +65 core, +296mhz mem: 1209 pts

Temps never exceeded about 67C, before repaste temps hit about 81C before I stopped the test (no oc)

See what I mean by a weird scenario? On my 290X I never had that happen, except for the opposite where overclocking the VRAM past a certain point would decrease the score, but never on the core oc combined with the memory. Yes, I know the X3230 is a bottleneck but that shouldn't affect it in GPU only benchmarks like Heaven or Valley, Timespy or Firestrike I could see but not these two. But it may be what you said, its probably because of newer games compressing and uncompressing textures being memory bound.
 

·
Registered
Joined
·
5,374 Posts
Heaven and Valley can run an unstable core OC, I wouldn't be surprised it that's causing the lower score. It takes longer but I'd try a hand full of games to make sure it's stable.
Sound's like you have this under control lol
 

·
Nvidia "shill"
Joined
·
1,429 Posts
Discussion Starter #10
Heaven and Valley can run an unstable core OC, I wouldn't be surprised it that's causing the lower score. It takes longer but I'd try a hand full of games to make sure it's stable.
Sound's like you have this under control lol
I did, I left one on for almost 24 hours unpaused and no crash happened.
 

·
stock...hahaha
Joined
·
1,622 Posts
One of the most useful tools that you can get for overclocking vram is OCCT

It will run a vram memory test that will show errors in real time. You can set memory run the OCCT test, f n errors increas the vram and test again. At a certain point, you will start to see the odd error. One error every second is fine, the card will error correct.

As you keep increasing the frequecy, the number of errors will start to exponentially increase. You want to set the overclock at the point where the errors are only infrequently appearing tn the OCCT test.

the Pascal performance drop at high memory frequencies is because the card is throwing many errors and the error correction is being overwhelmed
 

·
Registered
Joined
·
181 Posts
I did, I left one on for almost 24 hours unpaused and no crash happened.
I've noticed that just because it doesn't crash over 24hours doesn't mean it wasn't artifacting with misaligned polys or lost texture blinks during rendering. After you hit a final point I find it best to watch certain scenes very closely. I notice one of the scenes in Unigen heaven tend to expose a lot more artifacting than others. This would be the one it zoomed in to the inner hull of the flying ship, so just pause it there just before the fade out closest to the inner section and let it sit watching closely for artifacts for 15mins. At this point you should be pretty stable with GPU clocks.

On the other hand, this is also a good measure for the memory but I find that despite Unigen I had to knock back the memory a few notches in other games such as Mass Effect Andromeda as unigen did not stress the memory enough to expose any issues due to not heating up enough.

One of the most useful tools that you can get for overclocking vram is OCCT
Waitwait waaaaaaait.. OCCT does VRAM testing now? I need to look into that!
 

·
Nvidia "shill"
Joined
·
1,429 Posts
Discussion Starter #13
I mean I know how to overclock and stress test, I just need to know if overclocking the VRAM on this card would be more beneficial than the core because of the small memory bus. Also this card is very weird, I can increase the memory frequency all day long without adding voltage, but even adding +10mhz to the core means I need to add voltage, even with the memory clock untouched. Probably just bad silicon on this card i’m guessing.
 

·
9 Cans of Ravioli
Joined
·
19,918 Posts
Open Unigine Heaven. Start the benchmark. Pause the benchmark at a fairly high level load scene. The benchmark will still render the scene at whatever point you paused at.

Increase memory clock until you get artifacts or until performance drops.

I can max out the slider in Afterburner for my GTX 1060 (+1000) and not get any artifacts, but anything over ~ +600 brings performance down.
 

·
stock...hahaha
Joined
·
1,622 Posts
I mean I know how to overclock and stress test, I just need to know if overclocking the VRAM on this card would be more beneficial than the core because of the small memory bus. Also this card is very weird, I can increase the memory frequency all day long without adding voltage, but even adding +10mhz to the core means I need to add voltage, even with the memory clock untouched. Probably just bad silicon on this card i’m guessing.
Adding voltage only goes to vcore, the GPU frequency part od the chip. it does nothing to vdimm, the voltage that goes to memory.

There is no software control to adjust voltage to the memory, that would require hardware mods and the addition of a variable resistor.

Overclocking the vram can help with performance if the memory was bottlenecking the GPU performance. Try and see what happens to your benchmark results. The worst thing that can happen is make the PC crash and reboot.
 

·
Registered
Joined
·
5,374 Posts
Adding voltage only goes to vcore, the GPU frequency part od the chip. it does nothing to vdimm, the voltage that goes to memory.

There is no software control to adjust voltage to the memory, that would require hardware mods and the addition of a variable resistor.

Overclocking the vram can help with performance if the memory was bottlenecking the GPU performance. Try and see what happens to your benchmark results. The worst thing that can happen is make the PC crash and reboot.
Yep.
I believe some AMD cards have vram voltage control but I could be mistaken.
 

·
professional curmudgeon
Joined
·
10,402 Posts
personally i only OC vram when benching but found:

though no nvidia reference card has vdimm voltage control, adding a little to the core can help with the memory controller(s) on the chip.

valley responds really, really well to vram OCs but the clock speed achieved probably won't be stable in all games.

YMMV.
 

·
Nvidia "shill"
Joined
·
1,429 Posts
Discussion Starter #18
Yep.
I believe some AMD cards have vram voltage control but I could be mistaken.

Most AMD cards share the voltage with core and memory, or at least my 290X does. And yeah if I go over about +40mhz on the core my scores go down, and no i’m not thermal throttling, just is what it is.
 

·
Registered
Joined
·
5,374 Posts
I'm not sure about AMD cards. We are talking about a GTX 660
You realize I just agreed with you right? Does talking about one GPU mean that all others must be excluded?
We might be talking about a 660 but seeing as I know Twilight from other threads and clubs it was relevant and the main reason I posted anything here at all. Twilight has a 290/x which is why he might think that core voltage applies to ram as well which is why I said what I said.
 
1 - 20 of 22 Posts
Top