GeForce GTX 680, Part 2: SLI, 5760x1080, And Overclocking

Page 4 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
Status
Not open for further replies.
[citation][nom]Oldnconfuzed[/nom]Outstanding review, particularly given the degree of difficulty associated with max-fatigue!Looking for a couple opinions, Chris. (I'm predicating this on you having bought two of the AMD's as confirmation that you and/or the others in the lab like lots of horsepower)1) Given the choice between a real good 30" monitor and three 1080's, what's the consensus there?2) I've read that PCIE 3 has no material impact on performance given that PCIE 2 wasn't saturated. Do you concur or should we "stay tuned"?3) If you were to guess, will the much-speculated big-Kepler be for gamers or would it have an alternate target market?[/citation]
Not a problem, and thanks for the feedback!
1) I use three 1920x1200 Dells on my desk. I don't game on all three--just the center display--because until recently, AMD didn't make snapping windows to a screen an elegant process. As soon as I'm able to get some time, I will set this up. I *need* the three unique screens for productivity reasons, though.
2) PCIe 2.0 is still ample for any gaming-oriented workload. PCIe 3.0 will come into play when compute tasks start moving more data between graphics and system memory. Currently, the bus isn't a debilitating bottleneck.
3) My understanding is that the "big Kepler" will be substantially more complex than GK104. Then again, though, there must be a price to be paid in power, heat, and perhaps more relevant, dollars. That's just speculation of course, and I'm not sure what an economically-viable design (from a manufacturing perspective) would look like at this point.
 
[citation][nom]jaquith[/nom]Correct, PCIe 2.0 vs 3.0 has nearly no impact with the current resolutions and GPU's. You'll get the 0-3 FPS drop @ PCIe 2.0 with x8 lanes, and that's on a 5760x1200 render. On a typical 1920x1080 monitor, I really doubt any repeated tests will show anything. I do suspect once 4K monitor and GPUs that can properly drive them become more readily available then sure the differences will point to a PCIe 3.0 advantage. I've asked TH to do an article quite a while back, didn't I Chris?[/citation]
Yes you did David. *Chris pokes Thomas* This is your specialty, friend!
 
All of the ASUS LGA 2011 BIOS that I've seen allow you to switch between PCIe 2.0 and 3.0. It's a heck of a lot easier than the 'taping' method.

While you've got the set-up in place simply run a few tests. I say simply, I get 'simply' translates into a few hours... It would be a heck of good Reference Article that'll be used for quite sometime. :)
 
Nice review, guys! If You´re still thinking of what you could do for a third part, I would suggest micro stuttering evaluation and a multi-GPU scalability test. By a quick look at the results, NVidia seems to have improved over last generation, but who knows what a detailed test can show? Good to see we'll have a good fight between those two manufacturers.
 


Thanks so much for this article! It was great to see the main criticisms (OCing, 5760 res, AA) addressed. Two great cards, and both companies are clearly holding something in reserve (vRAM buffer for NV, clocks for AMD). I especially appreciated your clear statements about the variable 680 performance and thoughts on how aftermarket 680s will be able to enhance the cards.

One thing that made me curious is that NV was kind enough to send you two cards, but AMD was not. That seems to make your 7970 benches even more valuable. Did you look at the frequencies GPU Boost hit on your 680s? I was wondering if they are around the listed normal value of 1058 or if you ended up with nicer cards.

Edit: I'm not trying to imply anything about your/TH's methods-as I said, I greatly appreciated this review. The turnaround on addressing the concerns was fantastic! I'm just thinking about how PR/marketing departments (the people ultimately in charge of sending cards to reviewers) operate.
 
So what i got form this article is - if you don't intend to manually overclock your card, get NVidia. If you're not afraid of overclocking, get the 7970?
 
I still can't believe that AMD is keeping the base prices at the same level. If business mean anything to you drop the damm prices. 7970 at $449.99 would be a really, really great way to compete with 680...But hey if you think selling 1 or none for 549 is better than 10 for 449 good luck to ya.
 
[citation][nom]Hupiscratch[/nom]Nice review, guys! If You´re still thinking of what you could do for a third part, I would suggest micro stuttering evaluation and a multi-GPU scalability test. By a quick look at the results, NVidia seems to have improved over last generation, but who knows what a detailed test can show? Good to see we'll have a good fight between those two manufacturers.[/citation]

This is something that's making me hesitant about multi-GPU setups. Do I CF/SLI, or OC the heck out of a single GPU with a hefty watercooling rig?
 
[citation][nom]general lee[/nom]Both companies still need to work with their multi-gpu driver support. If I spend 1000$ on GPUs I'd expect them to work. At the moment there's too many games and settings where multi-gpu doesn't work. Add to that micro-stutters and increased noise and power consumption and multi-gpu doesn't seem like all that good of an idea.[/citation]

Only dual GPU setups ever have serious micro-stuttering, highly variable frame rates, etc where a single card setup doesn't. Remember, these phenomenon are not limited to dual GPU setups; single GPU setups can have it too. Three GPU setups tend to be equal to a single GPU setup with the same cards/GPUs when it comes to micro-stuttering and similar problems. Quad GPU setups tend to be even more stable/smooth than the single GPU setup, but often has driver problems that negate this advantage and the poor scaling of having more than two GPUs means you are paying a lot more money to get similar FPS, but smoother.

Performance scaling and micro-stuttering almost go hand in hand, but using two fast GPUs rather than two slow ones also decreases the problem. For example, dual GTX 580s has much less micro-stuttering than a dual GTX 550 TI setup does if all other variables are the same (same game, same CPU, etc).

Think of it this way. To have an average frame rate of 60FPS, that GPU needs to output 60 frames every second or one frame every 60th of a second. That's one frame every one and two thirds (1 2/3) of a millisecond. Lets say that two of the same GPU can hit about 110FPS. Now, you have a little over 1 and two thirds of a millisecond (slightly over one and four fifths) to make a frame, so you might think it should be smoother, but the reason for it not being smoother is simple... Two cards are pretty far apart and they need to exchange a lot of data very quickly. Sometimes, data isn't where it needs to be fast enough, and a frame takes longer to render than it should. This causes visual nuances on the screen. Three GPUs get 150FPS. Now, there's a third video card in the system that needs data too, but there is a lot more time to render a frame regardless of that because the FPS didn't go up a whole lot. Instead of 1 and 1 sixth of a millisecond, it has closer to 2 milliseconds. Despite there being a little more data shuffling, it's a lot more time, so it's a lot smoother than the two GPU setup.

Then you can add a fourth GPU that increases time to render a frame even more. However, for every GPU you add, there is a slightly higher latency. You'r mouse and such can only move once per frame per GPU and each GPU gets one frame at a time, so the mouse's movements only get rendered once every four frames. The sheer frame rate alleviates this somewhat, but it gets worse with every GPU added. I don't know at what point it becomes great enough to consider. Even with four GPUs, it shouldn't be much more than a 30% increase over a single GPU, so I don't think it's a problem, but I thought I might as well mention it.
 
I just love AMD. But, I however, can't deny that the Kelper GPUs are superior in almost every aspect to the 7000 series (except for compute performance).

I just don't feel comfortable buying Nvidia. especially when AMD is close behind. The fact that Nvidia stopped pushing for more compute performance in the desktop space when they were so focused on CUDA and GPGPU doesn't inspire trust. There's also their attitude towards hybrid PhysX. Call me a fanboy and thumb me down all you want but I am sticking with AMD at the current time. You can also call it a gut feeling that has served me very well over the years
 
[citation][nom]Energy96[/nom]Would have been nice to see how 580's in SLI compare to 680's in SLI. Afterall most people who would buy 2 680's likely already have 2 580's and would like to know if the upgrade is worth it or not.[/citation]
You are totally on target with this observation. I am one of those "most" people. I've noticed most sites won't post this type of review until 6-12 months after the fact. I think it might have something to do with the manufacturer's desires to not make the last gen cards seem obsolete just yet, so there might be some sort of written or unwritten agreement that says don't show last gen vs this gen as competing tech. Just a guess?
 
[citation][nom]shin0bi272[/nom]Im gonna guess its the AA but our friend "chip in a box" posted in the previous 680 review that in sli (even quad) hardwareinfo.com found at 5760x1080 in BF3 the crossfired 7970 was beating the 680... your review says differently though. Still this isnt bad for nvidia for a midrange card at least.Here's the pic he posted[/citation]

Chip in a Box's chart showed only that the VRAM of the 680 and/or multi-GPU scaling of the 680 falters after you put the third 680 in. Basically, this means that more than two 680s is probably not as good as more than two 7970s. That chart says that the 7970 retains great scaling even when adding a fourth 7970, something that has often been shown to be of little importance (for other video cards) for frame rate increases over three of the same GPUs. I think that Malmental said "VRAM. It also scales well." in response to that chart. Actually, I find it interesting that the fourth 7970 GPU retained excellent scaling. Usually, a fourth GPU doesn't make a big difference over three.

[citation][nom]raidenfox123[/nom]This is wrong, NVidia did not max there card out at all. Its running at a TDP of 175w under load where the max for a double 6pin is 225w..[/citation]

This very review shows that the GTX 680 doesn't get more performance from increasing memory clock frequency (in fact, it decreased performance) and that overclocking it's GPU didn't make much difference either. I'd say that it's fairly maxed out from the start.

Apparently, we can't go beyond 1125MHz on the shader cores because of driver problems or whatnot, but as Chris said, many 7970 owners found that the 7970 can easily reach 1300MHz. Judging by how well a 200MHz jump helped it in the next review, I daresay that the 7970 and GTX 680 could find themselves rough equals like the GTX 590 and Radeon 6990 (ironically, with similar performance to those two dual GPU cards too) do. The difference would then be the the 7970 has 50% more memory, a lot more memory bandwidth, a huge compute performance advantage, and probably has higher power usage.

A good review, but would have been better with some more information.
1. If other people are reaching more than 1125MHz, then regardless of how they did it, Tom's should have been able to find a way to do it too and then benchmark it's performance.

2. It would have been nice to see compute performance when overclocked. Understandably, Tom's probably didn't have enough time given the many things that needed testing, but it's still something I'd like to see, especially if the 7970 is put to around 1300MHz. Would also need to see power usage benchmarks at that speed too, so maybe another article can get devoted to this within the month?

3. It would have been nice to have multi-GPU scaling percentages, but like another said earlier, we can calculate that ourselves. Just from looking at the benchmarks without doing the math yet, it seems that the 680 scales extremely well, perhaps as well as or better than 6900.

4. At what point does the VRAM capacity difference become a problem? When is 2GB just not enough? The 7970 obviously has a huge advantage here, but just how important is it? 1080p often just inches above 1GB so I would assume that 2560x1440 and 2560x1600 would be nearing the limits of 2GB, but this review tells me that I'm wrong, so I have to ask when it becomes a problem.
 
[citation][nom]iknowhowtofixit[/nom]According to this: http://hwbot.org/hardware/videocard/radeon_hd_7970/The average OC on air is about 1200core/1700mem.[/citation]

Did you consider the possibility of that including people who didn't go past 1125MHz? All that number (1201MHz) tells us is that the average 7970 isn't forced above 1200MHz, but if a lot of people go for 1300MHz and a lot go for about 1100Mhz, well, that just shows you how high two sets of people are going, not how high it CAN go if you care to try.

[citation][nom]supall[/nom]So what i got form this article is - if you don't intend to manually overclock your card, get NVidia. If you're not afraid of overclocking, get the 7970?[/citation]

That's kinda the gist I'm getting, at least if the 7970 sees a price cut, because even at 1300MHz, it won't beat the 680 by much, if at all (although that should be enough to at least match it most of the time). Without a price cut, the 7970 is then more expensive for fairly similar performance (@~1300MHz), but it probably also uses more power, so it's really only a good buy if you do work like compute performance or play Metro 2033 and such. Otherwise, it might be like the Radeon 6990 versus the GTX 590... Trading blows, as high end graphics from different companies should. Nvidia did well with the GTX 680, but really should have left the DP to SP ratio at 1 to 8. With it being as low as it is, it takes two 680s to get an approximate match for even the GTX 580's DP performance.

Regardless of that, we also see that the 680 just didn't match 7900's SP compute performance anyway, so it would have needed something more like a 1 to 6 ratio just to match the 7970's DP performance. At that point, I could understand Nvidia wanting to protect the professional graphics cards, but even if they just left it at 1 to 8 and accepted a much less fatal loss to AMD here, it would have been a lot better and unless Nvidia intends to price their professional cards more than 7 or 8 times higher than the consumer versions, it shouldn't eat into professional card sales.

I simply don't understand the reasoning behind Nvidia having such a crippling DP to SP ratio. Even at 1 to 8, the DP price/performance of the professional cards should be higher unless Nvidia charges 8 times more money (or even more) than consumer cards for the same card in a professional version. If that, then at least keep the 1 to 11 ratio so it can at least keep up with the GTX 580 instead of being only half of it.
 
[citation][nom]youssef 2010[/nom]I just love AMD. But, I however, can't deny that the Kelper GPUs are superior in almost every aspect to the 7000 series (except for compute performance).I just don't feel comfortable buying Nvidia. especially when AMD is close behind. The fact that Nvidia stopped pushing for more compute performance in the desktop space when they were so focused on CUDA and GPGPU doesn't inspire trust. There's also their attitude towards hybrid PhysX. Call me a fanboy and thumb me down all you want but I am sticking with AMD at the current time. You can also call it a gut feeling that has served me very well over the years[/citation]

Nothing in this review proves that Kepler is superior in almost every way. Kepler is only beating a 7970 that is being limited by not being overclocked enough. Anyone who wanted to could hit 1300MHz on pretty much every single 7970 out there. From this article, it seems as though doing this disables some features because there isn't proper driver support for decent overclocks right now (-1 for that AMD, forcing us to not even use your overpriced card to it's true limits).

In fact, the Tahiti seems superior to the GK104 in more ways than the GK104 is to the Tahiti. Tahiti has a lot more memory bandwidth and capacity and seems to have greater overclocking headroom. However, like some other poster said (can't remember the name, not looking for it), changing the target TDP of the GTX 680 doesn't seem to be the best method for overclocking it. Apparently, it is better to just manually increase the frequencies like it is with other video cards. Kepler does seem to be more energy efficient at load, but we still don't know enough to make well educated conclusions just yet. Perhaps a third review will answer most of the questions that still remain.

Kepler has some advantages, but 7900 has it's own too and I have no intention of downplaying either side's advantages and disadvantages.

Otherwise, I agree with you.
 
The WOW 5760x1080 Benchmark is reading that the 7970 is faster w/ 8xAA than no AA. I think this is a Typo and the numbers should be switched.
 
[citation][nom]bwb34[/nom]The WOW 5760x1080 Benchmark is reading that the 7970 is faster w/ 8xAA than no AA. I think this is a Typo and the numbers should be switched.[/citation]

Might not be a typo. The article tells us that WoW wasn't being very cooperative with the 7970 CF setup.
 
To me the difference in performance, not really being all that much either way, is quite substantial when comparing them at a design/ architectural standpoint. To use an analogy to exemplify my idea, let's say that the gtx-680 is a Mercedes and the 7970 is a Ferrari. Negating the price difference between the two and simply looking at performance, the specs for the Ferrari should always outperform the specs for the Mercedes. The Ferrari/ 7970 having 4.3 billion transistors, a 384 bit bus, and 3GB of memory and the Mercedes/ gtx-680 having 3.5 billion transistors, 256 bit bus, and 2GB of memory. Certainly the Ferrari/ 7970 is more of a high-end beast than the Mercedes/ gtx-680 which makes you wonder how a lower performing product can keep a marginal gain on the competition, especially when you figure that the 7970 has had three months market time to solve consumer driver issues which the 680 is basically tuned to reference drivers. To use the car analogy again, the Ferrari most likely cost a lot more to develope and produce than the simpler Mercedes since the Ferrari is a lot more complex in design, therefore the profit margins for the Ferrari will be a lot less than the Mercedes. Literally meaning that Ferrari can only sell cars at a limited and high price point where as at the moment since the Mercedes/ gtx-680 can outperform the Ferrari/ 7970 and Mercedes can upsale their car at the same price point for a much higher profit therefore creating a distinct edge in price competition. The analogy used is only to illustrate the spec differences between the two cards, in the real world a Ferrari will generaly be far more expensive and always outperform a mercedes, I only use these cars as an example because they are both high-end luxury cars that really don't compete in the real world but in the case of the gtx-680 and 7970, one card should vastly outperform the other but doesn't. I am by no means an expert on graphics cards and this is just my way of interpreting the results of endless hours of professional tests and benchmarks. After initially reading the review I had a mental picture of a guy on a moped with a bumper sticker that said "eat my dust" with a guy on a Harley right behind him.
 


What I was more concerned with was that here at Toms they had the 680 in sli beating the 7970 in crossfire (dual card) and the other review has the 7970 ahead in dual card at the same res. But other than that yeah.
 
[citation][nom]orwellswift[/nom]To me the difference in performance, not really being all that much either way, is quite substantial when comparing them at a design/ architectural standpoint. To use an analogy to exemplify my idea, let's say that the gtx-680 is a Mercedes and the 7970 is a Ferrari. Negating the price difference between the two and simply looking at performance, the specs for the Ferrari should always outperform the specs for the Mercedes. The Ferrari/ 7970 having 4.3 billion transistors, a 384 bit bus, and 3GB of memory and the Mercedes/ gtx-680 having 3.5 billion transistors, 256 bit bus, and 2GB of memory. Certainly the Ferrari/ 7970 is more of a high-end beast than the Mercedes/ gtx-680 which makes you wonder how a lower performing product can keep a marginal gain on the competition, especially when you figure that the 7970 has had three months market time to solve consumer driver issues which the 680 is basically tuned to reference drivers. To use the car analogy again, the Ferrari most likely cost a lot more to develope and produce than the simpler Mercedes since the Ferrari is a lot more complex in design, therefore the profit margins for the Ferrari will be a lot less than the Mercedes. Literally meaning that Ferrari can only sell cars at a limited and high price point where as at the moment since the Mercedes/ gtx-680 can outperform the Ferrari/ 7970 and Mercedes can upsale their car at the same price point for a much higher profit therefore creating a distinct edge in price competition. The analogy used is only to illustrate the spec differences between the two cards, in the real world a Ferrari will generaly be far more expensive and always outperform a mercedes, I only use these cars as an example because they are both high-end luxury cars that really don't compete in the real world but in the case of the gtx-680 and 7970, one card should vastly outperform the other but doesn't. I am by no means an expert on graphics cards and this is just my way of interpreting the results of endless hours of professional tests and benchmarks. After initially reading the review I had a mental picture of a guy on a moped with a bumper sticker that said "eat my dust" with a guy on a Harley right behind him.[/citation]

Sorry, but it doesn't work that way. How many transistors a chip has does not define performance (it never has) exactly, how wide it's memory interface does not define performance exactly, how much memory capacity they have does not define performance exactly. There are huge differences and factors that you ignored because you don't know what you're talking about. I'm no expert, but I think I can try to clarify a little more.

Kepler seems to have greater gaming performance per core than GCN if a 1536 core chip clocked at between about 1GHz and 1.056GHz beats a 2040 core chip clocked at 925MHz. Kepler is obviously inferior to GCN for SP floating point performance (DP is artificially lowered by Nvidia, not low because of problems with the chip so I won't go into it right now) because even though the 680 is a higher performing card at stock than a stock 7970 for most games, the 7970 has a huge lead in SP computational performance. However, the SP performance advantage of the 7970 seems to be rather similar to it's core count advantage, so the Kepler and GCN cores probably have similar computational performance.

The 680 has less efficient AA performance for most of the titles than the 7970. This is likely because of it's memory bandwidth bottleneck. Of course, it's performance lead is enough to not let the 7970 retake the crown at stock frequencies, but it shows promise for the 7970 when overclocked without being held back by the current version of Catalyst.

The 7970 has a big VRAM capacity advantage of 50% too. That usually won't matter for most setups until you hit things like Eyefinity resolutions and such, but it may become more important in the future, especially with future-proofing. If two 680s or 7970s are necessary for 2560x1600 in some new games in a few years, well then you can bet that a mere 2GB of memory might become a problem. There's a reason that AMD mostly uses 2GB for their 6900/6800/7800 only, not their even faster 7900 cards. However, Nvidia seems to really like skimping on VRAM capacity as you can see them do it with multiple cards, not just the 680. The 570 only has a mere 1.25GB, limiting going far over even 1080p. The GTX 580 isn't much better with only 1.5GB of VRAM, but plenty 580s have 3GB so it's a little different whereas only one or two 570s has 2.5GB.

The 7970, when overclocked past Catalyst's limits to about 1300MHz, will be roughly similar to the 680. All the 7970 needs to compete with the 680 in performance is for Catalyst to stop limiting overclocks to 1125MHz. Better drivers for both cards will come, then we will see if either card is truly the better. As of right now, it seems like the 680 is more power efficient. However, we have yet to see the 7970 here on Tom's compared against the 680 when both are truly overclocked to their reasonable limits. The 7970 obviously has a lot more performance headroom, but the 680 seems to have hit a cliff in performance almost out of the box.
 
[citation][nom]shin0bi272[/nom]What I was more concerned with was that here at Toms they had the 680 in sli beating the 7970 in crossfire (dual card) and the other review has the 7970 ahead in dual card at the same res. But other than that yeah.[/citation]

Looking at the test titles, you notice that the exact test being run in both Chip's picture and this article's picture are slightly different tests. Maybe that's why? Even in this review, the 7970 and 7970 CF were right behind the GTX 680, maybe in the test shown by Chip, the 7970 pulled ahead with CF of the SLI setup. It would be nice to see the review where Chip's picture came from to find out if it's actually fake or not because that possibility is still on the table.
 
[citation][nom]blazorthon[/nom]Looking at the test titles, you notice that the exact test being run in both Chip's picture and this article's picture are slightly different tests. Maybe that's why? Even in this review, the 7970 and 7970 CF were right behind the GTX 680, maybe in the test shown by Chip, the 7970 pulled ahead with CF of the SLI setup. It would be nice to see the review where Chip's picture came from to find out if it's actually fake or not because that possibility is still on the table.[/citation]

yeah I think in the test chip posted it had 4x msaa on and in the toms one it was using no aa... still damned good for 256bit mem interface. Cant wait to see what happens when they release one with 384bit.
 
[citation][nom]shin0bi272[/nom]yeah I think in the test chip posted it had 4x msaa on and in the toms one it was using no aa... still damned good for 256bit mem interface. Cant wait to see what happens when they release one with 384bit.[/citation]

Tom's benchmarked both no AA and with 4x MSAA so that's not the difference.
 
Off the bat, I'm more so an AMD fanboy, I'm always pleased when they come out on top and root for them deep inside (the whole underdog thing). When I build rigs to sell, I use Radeon more than Nvidia. It was awesom seeing the 7970s kicking behind to the 580 gtx when it came out, however that's a next gen card competing with a last gen card...victory was short lived

Ultimately, when it comes to loyalty, I take the realist approach and personally go with "what is faster". Hence, I'm a proud owner of a PNY 680 GTX (their website still has them in stock - ordered and it came in 3 days - go figure...). Performance is rock-solid and I was surprised on how cool and quiet my card runs, even under load. It's not as fast as the 2 470 GTX's I had in SLI but it's a single card solution and I've got a second one the way for kick a@$ performance (and kick a@$ credit card bill). I will say that AMD is close to closing the gap with Nvidia but for FPS #'s and bragging rights, Nvidia is the winner. Of course AMD seems to do well with computing power - not sure if that's more of a selling point to overtake Nvidia among gamers and custom PC botiques.
 
I find there's something really fishy about this 680GTX card.
I wonder if the image quality is exactly the same from one camp to another even with no AA.

Plus, IMHO FXAA is total garbage. It does miracles for jaggies but seems to compromise the sharpness of textures in general. May as well look at your monitor through a foggy window and retain your FPS than activate useless FXAA.

 
Status
Not open for further replies.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom