Sign in with
Sign up | Sign in
Your question

670 2GB, 670 4GB, or 2 x 660 Ti 3GB in SLI ?

Last response: in Graphics & Displays
Share
November 29, 2012 4:00:51 PM

Currently on a 1920x1200 set up with the potential to go up to 2560x1600 in the next year or two.

On the single card front I am deciding between the EVGA 670 FTW 2GB and 4 GB version.

An alternative is getting 2 EVGA 660Ti FTW+ 3GB version and run them in SLI.

When running BF3 (or even Crysis Warhead) everything to the max, I see the FTW 2GB already using up 1.8GB of video memory. I seriously wonder for Crysis 3 or Arma 3 they can use more than 2GB of VRAM even for 1920x1200.

I know very well either version of the 670 doesn't do 2560x1600 very well for the highest level of detail, while putting 2 660Ti in SLI does.

I have a 850HX supplying the juice to a i7-2600K Z68 SLI-capable system so power draw is a non issue. Heat is the only thing that I can't stand. The 7970 GHz edition 3GB was ruled out as the heat it generate is not justified by the performance. A whole 100Watt extra over the 670 for perhaps 10-20% more performance. The 660Ti SLI might generate as much beat but the 1.5x to 1.8x performance over single 670 seems more worthy.

What would you pick?
a b U Graphics card
November 29, 2012 4:10:43 PM

pxl9190 said:
Currently on a 1920x1200 set up with the potential to go up to 2560x1600 in the next year or two.

On the single card front I am deciding between the EVGA 670 FTW 2GB and 4 GB version.

An alternative is getting 2 EVGA 660Ti FTW+ 3GB version and run them in SLI.

When running BF3 (or even Crysis Warhead) everything to the max, I see the FTW 2GB already using up 1.8GB of video memory. I seriously wonder for Crysis 3 or Arma 3 they can use more than 2GB of VRAM even for 1920x1200.

I know very well either version of the 670 doesn't do 2560x1600 very well for the highest level of detail, while putting 2 660Ti in SLI does.

I have a 850HX supplying the juice to a i7-2600K Z68 SLI-capable system so power draw is a non issue. Heat is the only thing that I can't stand. The 7970 GHz edition 3GB was ruled out as the heat it generate is not justified by the performance. A whole 100Watt extra over the 670 for perhaps 10-20% more performance. The 660Ti SLI might generate as much beat but the 1.5x to 1.8x performance over single 670 seems more worthy.

What would you pick?

Let's get our facts straight (about power consumption): http://www.techspot.com/review/603-best-graphics-cards/...

@ Needed RAM in GPU: Well i was just reading an article about a new demanding game title in guru3d. They made a frame buffer usage test. Maybe it helps: http://www.guru3d.com/articles_pages/hitman_absolution_...

You need 1.5GB+ graphics memory @ full HD to max out Hitman: Absolution. So, i have no comment on that, you can draw your conclusion :) 
November 29, 2012 4:25:15 PM

Get 670 4gb and get another one for SLI when you actually get that monitor u want.
Related resources
a b U Graphics card
November 30, 2012 8:51:18 AM

NO GTX 660 ti is HORRIBLE for anything above 1080p with its crippled 192 bit even in SLI it will struggle.

You are much better of with an AMD card at that rez, even if they generate "too much heat"? (I lol'ed at that one). From 1600p onwards, AMD starts to show a much better performance than Nvidia cards, with its 3 GB memory and 384 bit bandwidth. So even if they don't look good now, they will benefit in the future for SLI and when you get the monitor you want

However if your an Nvidia fan, the 4 GB GTX 670 will probably be the best option, although I don't think it will be easy to get another 4GB 670 in the future when you want to SLI as that is not the stock memory and is rare to find.
November 30, 2012 11:09:04 AM

Ravyu said:
NO GTX 660 ti is HORRIBLE for anything above 1080p with its crippled 192 bit even in SLI it will struggle.

You are much better of with an AMD card at that rez, even if they generate "too much heat"? (I lol'ed at that one). From 1600p onwards, AMD starts to show a much better performance than Nvidia cards, with its 3 GB memory and 384 bit bandwidth. So even if they don't look good now, they will benefit in the future for SLI and when you get the monitor you want

However if your an Nvidia fan, the 4 GB GTX 670 will probably be the best option, although I don't think it will be easy to get another 4GB 670 in the future when you want to SLI as that is not the stock memory and is rare to find.


Yeah going w/a 7950 or corssfire 7870s is a good option for higher resolutions the memory bus on the 660 sucks at 192 bit.
a c 291 U Graphics card
November 30, 2012 11:29:16 AM

I'd go with SLI GTX 660 Ti x2. Looks like memory doesn't hurt too much:





Memory will not be limit even at 2560x1600. I'd like to tell you this: just because it uses 1.8 GB of VRAM when it has available, doesn't mean it will lag if you have less VRAM. Game engines have special optimizations so the memory is swapped seamlessly, and even when VRAM is maxed out, you can still enjoy 60 fps easily (like in Hitman Absolution, I play on mix of High-ultra settings on my GTX 560 Ti 1 GB and even though VRAM gets capped, I still get 50-60 fps all the time without any fps drops).

As for people saying AMD will be better for the future: for the past 10 years memory bus width requirements didn't grow: even the old FX 5900 had 256 bit bus, and it was released 9 years ago. It will not change in the future, and if it will, it will not be sudden.
a b U Graphics card
November 30, 2012 11:48:53 AM

Sunius said:
As for people saying AMD will be better for the future: for the past 10 years memory bus width requirements didn't grow: even the old FX 5900 had 256 bit bus, and it was released 9 years ago. It will not change in the future, and if it will, it will not be sudden.
True but Kepler's core architecture is much different than old FX's and we see many scenarios where memory bandwidth is a limiting factor for almost any Kepler card. For example throw some heavy MSAA in any given game and watch what will happen with Keplers. If a GTX 670/680 had 384bit interface, they would be MUCH ahead (maybe +30%/40%) of ANY AMD out there. What i mean:

http://www.tomshardware.com/reviews/geforce-gtx-660-ti-...
a b U Graphics card
November 30, 2012 1:13:36 PM

Just to clear up this old bus width argument, check out the second post:

http://www.tomshardware.co.uk/forum/378279-15-gtx660-78...

7900-series cards deliver a lot more bandwidth, as do GTX670/680. For comparisons between GTX660/660 Ti and 7800-series cards, it's irrelevant. But as has already been said, stop worrying about specs and just look at the results :-)
a b U Graphics card
November 30, 2012 2:12:10 PM

sam_p_lay said:
Just to clear up this old bus width argument, check out the second post:

http://www.tomshardware.co.uk/forum/378279-15-gtx660-78...

7900-series cards deliver a lot more bandwidth, as do GTX670/680. For comparisons between GTX660/660 Ti and 7800-series cards, it's irrelevant. But as has already been said, stop worrying about specs and just look at the results :-)

No offense but we are already looking at the results, that's what i am saying also
a b U Graphics card
November 30, 2012 2:18:10 PM

burritobob said:
Yeah going w/a 7950 or corssfire 7870s is a good option for higher resolutions the memory bus on the 660 sucks at 192 bit.


^ That's what I'm responding to, and Ravyu's 'with its crippled 192 bit even in SLI it will struggle'. It's important people are aware that this tired old 'crippled bus width' argument is totally false.

EDIT: Also... you maybe missed the 'as has already been said' in my post. That's effectively a +1 to the people who said it - it's not stating that nobody has mentioned it.
a b U Graphics card
November 30, 2012 4:31:58 PM

sam_p_lay said:
It's important people are aware that this tired old 'crippled bus width' argument is totally false.

I really mean no offense but i must clearly state that i DRAW MY OWN conclusion that SOMETHING hurts Nvidia Kepler cards badly. IT LOOKS LIKE it is about memory bandwidth as stated in this following article which was prepared by a respectable Tom's Hardware reviewer (i linked above, i link again below). I can link to other reviews from respectrable sites doing similar tests and getting similar results. SO it doesn't look like AN OLD argument because Kepler is the newest generation GPU architecture of Nvidia, sold in the market for real money. What am i emphasizing?: If you are using MSAA heavily, Kepler is NOT for you. If you are doing bandwidth requiring stuff, Kepler is NOT for you. I am IN NO WAY a fan of AMD nor Nvidia by the way.

http://www.tomshardware.com/reviews/geforce-gtx-660-ti-...
a b U Graphics card
November 30, 2012 4:37:39 PM

That article is a lot of the reason people keep going on about bus width. It's an old argument in that it's been said again and again since the card's release. When several people are saying it each day in the forums, it gets old fast. Did you even read the explanation of bus width vs bandwidth I posted?

tl;dr? Bus width is no more important than memory clock frequencies - they're both factors of bandwidth, which is what counts. GTX660 and GTX660 Ti shift 144GB/s. 7850 and 7870 are 153GB/s. Not exactly the huge deal it's made out to be.

Please ease off on the caps lock.
a b U Graphics card
November 30, 2012 4:56:23 PM

I don't agree you on that. I really believe a high bandwidth plays a huge role in how today's cards perform, memory speed at the 2nd place IMO. And what i mean is that it can be a timely old argument but it still applies. Are Keplers good? Hell yes. I'd pick a GTX 670 if i could today, but i would do it knowing what i get/what i am ready to sacrifice. So any buyer asking for advise in these forums needs to compare different arguments and make their decision based on their own needs/conclusions. They must know/get informed about what they pay for and what they get back. We share our knowledge here, even if we agree or not. So thank you for sharing your point of view too
a b U Graphics card
November 30, 2012 5:05:57 PM

You haven't read what I've written properly. I said bandwidth does matter, it's bus width that doesn't (at least not taken in isolation of clock freqs). I said what isn't a huge deal is the amount of difference - 6% in this case. You should take a look at this, you might learn something:

http://www.tomshardware.co.uk/forum/378279-15-gtx660-78...

What's written in that post is fact, not opinion. There's nothing subjective there. So if anything is inaccurate, feel free to point it out. A lot of this stuff is hard fact backed up by numbers - nothing to do with opinion. Where knowledge is inaccurate/incomplete, it's important to educate people so they make those informed decisions you mention. If they're told that bus width is important (as opposed to bandwidth or just the end result - frames/second) then they've been misled.

EDIT: If you read that, you'll learn memory speed is a factor of bandwidth - it's not a separate aspect. Multiply memory clocks with bus width (remembering the 8 bit/byte division) by transfers per clock cycle and you get bandwidth.

This stuff is just maths - 10 *10 = 100. 5 * 20 = 100. The result in both cases is 100 and it doesn't matter how you get there. 100 is the bandwidth and that's what impacts performance. By ~6% in this case.
a b U Graphics card
November 30, 2012 5:48:02 PM

sam_p_lay said:
You haven't read what I've written properly. I said bandwidth does matter, it's bus width that doesn't (at least not taken in isolation of clock freqs). I said what isn't a huge deal is the amount of difference - 6% in this case. You should take a look at this, you might learn something:

http://www.tomshardware.co.uk/forum/378279-15-gtx660-78...

What's written in that post is fact, not opinion. There's nothing subjective there. So if anything is inaccurate, feel free to point it out. A lot of this stuff is hard fact backed up by numbers - nothing to do with opinion. Where knowledge is inaccurate/incomplete, it's important to educate people so they make those informed decisions you mention. If they're told that bus width is important (as opposed to bandwidth or just the end result - frames/second) then they've been misled.

EDIT: If you read that, you'll learn memory speed is a factor of bandwidth - it's not a separate aspect. Multiply memory clocks with bus width (remembering the 8 bit/byte division) by transfers per clock cycle and you get bandwidth.

This stuff is just maths - 10 *10 = 100. 5 * 20 = 100. The result in both cases is 100 and it doesn't matter how you get there. 100 is the bandwidth and that's what impacts performance. By ~6% in this case.

Ok then you can take a look at this sentence (from that review that i linked):
"We aren't saying that anyone else ran their benchmarks on Nvidia's GeForce GTX 660 Ti incorrectly. Really, the card that wins depends on games, settings, and resolutions. This card isn’t a good choice for less demanding titles, but it does make a strong showing when a lot of GPU performance and, relatively speaking, not a lot of memory bandwidth are needed. But this exposes the card's big issue. Nvidia's GeForce GTX 660 Ti isn’t really a premium card. It would need to perform better at higher-end settings to satisfy the folks shopping for a less expensive alternative to the GTX 670. If you're really only looking for a middle-of-the-road card to game at mainstream resolutions and modest settings, the less expensive Radeon HD 7870 is ample."
By Igor Wallossek - Toms Hardware.

If it is not bandwidth, maybe you can explain what causes a GTX 66Ti (or GTX 670 in this case) to get less performance when heavy MSAA is involved. If it is not bandwidth how do we call it? If you are more comfortable by calling it bus width let it be so lol
a c 291 U Graphics card
November 30, 2012 6:01:01 PM

Bus width and bandwidth are two different things...
a b U Graphics card
November 30, 2012 6:03:07 PM

You continue to miss the point... I'm just not getting through to you. And you're clearly not reading what I write properly because it has nothing to do with what I prefer to call it.

bus width x memory clock frequency x transfers per clock cycle = bandwidth

That is an equation. It has nothing to do with nomenclature or opinions or anything else. It's hard fact. You can say 'I genuinely believe 2 + 2 = 5' but that doesn't make it so. You can say 'No offence - I respect your opinion that 2 + 2 = 4 but I respectfully disagree'. We're not talking about opinion here. 144 is a number. 153 is a number. Those are the numbers that represent bandwidth of the two pairs of GPUs. That's a 6% difference.

Now what the THG review is getting at is that the GTX660 Ti doesn't deliver the same memory bandwidth as GTX670/HD7970. And that's correct. Those GPUs offer hugely superior bandwidth. BUT they're in a different price range. Compared to the 7800-series cards, there is no significant bandwidth difference.

As for what is responsible for the difference in heavy MSAA performance, I don't know, and I don't care. Anything I say on it (or anyone else outside of nVidia) would be pure speculation. If you think what I've said about bandwidth is inaccurate, go ahead and get in touch with Chris Angelini. His email is publicly accessible on the site and he's a good guy - you'll get a detailed and helpful response confirming what I've told you.
a b U Graphics card
November 30, 2012 6:16:23 PM

Man i understood what you mean. Sorry if i call the "concept" wrongly. I called the term the way how it is commonly used. My point is that under certain circumstances Kepler cards get poor performance relative to Radeons, MSAA and very high resolutions being 2 of those, but Radeons can't do phsyx etc etc That's what i mean, everybody must know what is good and what is bad about what they buy. That is what i am struggling to tell. So i don't give a heck how this is called, call it an "elephant" or "monkey", that is not my point. My point is the repeatable result we get by doing the same exact test. I am not talking about technical stuff here, no babbling on technical terms, i am also talking about the cold hard fact.
a b U Graphics card
November 30, 2012 7:08:37 PM

Alright well I think most readers will understand what I'm talking about anyway. What I posted wasn't that technical anyway - fairly basic concept but worth understanding.
a b U Graphics card
November 30, 2012 7:22:14 PM

sam_p_lay said:
Alright well I think most readers will understand what I'm talking about anyway. What I posted wasn't that technical anyway - fairly basic concept but worth understanding.

No, really i understand what you mean. You say this MSAA situation can't be the result of lower bus width (here you go :) ) because at the end both radeons and keplers have similar/comparable bandwidth, resulting from different memory clock speeds x bus width = bandwidth. I knew what you mean from the begining. But this is commonly adressed/called "bandwidth limitation" on the web. I mean as the calling, how they name it. They somehow relate this result to the lower bus width of Nvidia cards, and i based my assumptions on these
a b U Graphics card
November 30, 2012 7:36:00 PM

That's cool :-) Yeah the 7900s and GTX670/680 have massively superior bandwidth (192 and 240/264 respectively if you're interested) but they are high-end, I just wanted to clarify the situation for the mid-range stuff, since SLI GTX660 Tis are being considered, and are at no significant bandwidth disadvantage to the 7870 (which is their competition).

Anyway, SLI GTX660 Ti (or SLI GTX660) will deliver massively superior framerates to a GTX670, but at the drawback of less consistent framerates and potential issues with performance and/or visual glitches in certain games that aren't initially optimised for multi-GPU setups. You have to then be constantly updating drivers and messing around with SLI profiles for games... maybe it's worth it, but I'd personally grab a GTX670 for more consistent, problem-free performance.

And as for memory quantity, the post about future availability of 4GB models is a very good point and something I'd always consider too. Same goes for non-stock clock speeds (i.e. with factory-overclocked cards) - they'll need to match, with means either adjusting clock speeds if you can't get a matching card in the future (and voiding warranty) or sticking with stock clock models in the first place to avoid complications down the line. I'd personally take a 2GB model.

Finally, the Ti really doesn't add a lot over a non-Ti GTX660, so like others have pointed out, isn't really a great buy relative to the GTX660. I'd skip straight to GTX670 if you want something genuinely faster.

EDIT: I'm only calling those cards 'mid-range' because their price is mid-range - performance really isn't miles off the top-end stuff. They really do offer much better value for money than GTX670/680 and HD7970.
November 30, 2012 7:52:59 PM

670 once again.
a b U Graphics card
November 30, 2012 8:03:16 PM

Now that GTX 670 and HD 7970 is similarly priced, i say 7970
a b U Graphics card
November 30, 2012 8:29:22 PM

I think price depends on country - are we talking about £, $, € or ru? In the UK you're looking at £270+ for a GTX670 vs £300+ for a 7970. In other countries, those prices might be closer.
December 2, 2012 5:59:42 AM

Sunius said:
I'd go with SLI GTX 660 Ti x2. Looks like memory doesn't hurt too much:

http://media.bestofmicro.com/9/P/348973/original/Multicard%20Battlefield3.png

http://media.bestofmicro.com/9/S/348976/original/Multicard%20Maxpayne.png

Memory will not be limit even at 2560x1600. I'd like to tell you this: just because it uses 1.8 GB of VRAM when it has available, doesn't mean it will lag if you have less VRAM. Game engines have special optimizations so the memory is swapped seamlessly, and even when VRAM is maxed out, you can still enjoy 60 fps easily (like in Hitman Absolution, I play on mix of High-ultra settings on my GTX 560 Ti 1 GB and even though VRAM gets capped, I still get 50-60 fps all the time without any fps drops).

As for people saying AMD will be better for the future: for the past 10 years memory bus width requirements didn't grow: even the old FX 5900 had 256 bit bus, and it was released 9 years ago. It will not change in the future, and if it will, it will not be sudden.


Thanks for posting this cause there alot of BS about the 660 ti 2gb card that is so untrue..i personally have the evga 660 ti in SLI and i can max every game out no problem with absolutely no heat issues and i run them on a evga nex750 psu with 8 fans hooked up, so power consumption is fantastic with these cards. Im running a solid 90-95fps in bf3 with no micro stutter and 200+fps in black ops 2. The 660 ti sli'd is the absolute best bang for the buck and the 192 bit doesnt effect it at all. Plus you get some good free games, assasins creed 3 and borderlands 2 all at an afordable price.
Anonymous
March 20, 2013 4:51:09 AM

I think the reason why nobody can agree on what brand has better GPU's is because there aren't 2 cards from both brands that have exactly the same specs and price. It seems like AMD looked at Nvidia's line up and basically inserted a card in between each tier of Nvidia cards. If you start at the top and work your way down the entire lineup of both brands, it will probably look like Nvidia > AMD > Nvidia > AMD > Nvidia > AMD > Nvidia > AMD etc. There aren't really two cards that overlap to make a good comparison. And none of the cards overlap in price either. The closest AMD card is always cheaper. So when people are shopping, AMD is betting on people going for their slightly less expensive but similarly spec'd card or hoping that they will go for the AMD card which is only slightly more expensive than the Nvidia but performs almost as well as the next step up Nvidia. This is true for the Intel vs AMD CPU/APU shopping choice dilemma as well. AMD is the bargain brand and many elitist snobs usually pass on their products because of this. But, the fact is: AMD is never the better product. AMD is never the fastest product. They always seem to be the 2nd, 3rd or 4th fastest for the last few generations of CPUs and GPUs. They seem to care more about being the more affordable choice than the better choice. Everything about Intel and Nvidia is better than a similar AMD product; performance, quality, efficiency, temperature - except for price. You do, however, get more "bang for your buck" with AMD, which is nice. It is like a Corvette ZR1 vs a Ferrari 458. The Chevy costs about half as much and even has a few more horsepower and torque. But the Ferrari destroys it in every way except price and a few things on the spec sheet. The Chevy is the better bang for your buck, but the Ferrari will beat it in a straight line and around a track every single time and do it while using less gas.
a b U Graphics card
March 20, 2013 7:10:17 AM

Anonymous said:
I think the reason why nobody can agree on what brand has better GPU's is because there aren't 2 cards from both brands that have exactly the same specs and price. It seems like AMD looked at Nvidia's line up and basically inserted a card in between each tier of Nvidia cards. If you start at the top and work your way down the entire lineup of both brands, it will probably look like Nvidia > AMD > Nvidia > AMD > Nvidia > AMD > Nvidia > AMD etc. There aren't really two cards that overlap to make a good comparison. And none of the cards overlap in price either. The closest AMD card is always cheaper. So when people are shopping, AMD is betting on people going for their slightly less expensive but similarly spec'd card or hoping that they will go for the AMD card which is only slightly more expensive than the Nvidia but performs almost as well as the next step up Nvidia. This is true for the Intel vs AMD CPU/APU shopping choice dilemma as well. AMD is the bargain brand and many elitist snobs usually pass on their products because of this. But, the fact is: AMD is never the better product. AMD is never the fastest product. They always seem to be the 2nd, 3rd or 4th fastest for the last few generations of CPUs and GPUs. They seem to care more about being the more affordable choice than the better choice. Everything about Intel and Nvidia is better than a similar AMD product; performance, quality, efficiency, temperature - except for price. You do, however, get more "bang for your buck" with AMD, which is nice. It is like a Corvette ZR1 vs a Ferrari 458. The Chevy costs about half as much and even has a few more horsepower and torque. But the Ferrari destroys it in every way except price and a few things on the spec sheet. The Chevy is the better bang for your buck, but the Ferrari will beat it in a straight line and around a track every single time and do it while using less gas.

This isn't true for GPUs because both Nvidia and AMD offer comparable quality, performance, efficiency, features etc. Not possible to draw an obvious conclusion like in CPUs. What you say is true for CPUs though
April 3, 2013 2:28:34 PM

Can you run a gtx 660 ti 4GB in sli with a gtx 660 ti 2GB. and if so can you make the 4GB the Domitian one so you are still utilizing the 4gigs of ram.
Anonymous
April 25, 2013 4:58:24 PM

trickedout said:
Can you run a gtx 660 ti 4GB in sli with a gtx 660 ti 2GB. and if so can you make the 4GB the Domitian one so you are still utilizing the 4gigs of ram.


Yes, you can run two cards with the same GPU but with different VRAM sizes. The 4GB card will only use 2GB of its RAM because it will be limited by the 2GB card. The two cards each do every other frame, from what I've read, and each use their own RAM since the SLI bridge can't handle that high of a transfer rate. Since the game is limited to how ever high the quality settings can be within 2GB of VRAM, the other card will be limited to 2GB as well.

Hopefully, one day, they will come out with something similar to SLI that will allow two or more cards to share VRAM. I'm guessing that it will require a bridge that looks somewhat like a PCI-E slot with a bunch of gold plated teeth, only on top. I am also guessing that the current SLI bridge only serves to tell the PC that the two cards are working together. All the data, in and out, is still being transferred through the PCI-E slots and the SLI bridge is just a sort of timing device. The reason why SLI doubles performance, though, is because each GPU is only rendering half the frames in the same frame rate.
!