Nvidia GeForce GTX Titan 3-way/4-way SLI review incl. 5760x1080 and frametimes

[Final8ty]

How do two, three and four GTX Titans perform?

By Koen Crijns, Saturday March 2, 2013 12:58 AM

Introduction

Yesterday we published our comprehensive review of the new Nividia GeForce GTX Titan graphics card, but it lacked one important item. How do two, three and four Titans perform together? We tested all SLI combinations, including 5760x1080 and frametime tests, and compared the results to GeForce GTX 690 Quad-SLI.

Nvidia itself only mentions SLI and 3-way SLI, and indicated that 4-way SLI would not be possible with the GTX Titan. However, the Nvidia-sponsored overclocker Kingpin has already posted 3DMark scores with four Titan cards on HWbot.org. We of course had to try our own 4-way SLI configuration, and we're happy to report that it's definitely possible.

And despite the rather steep pricetag of $1,000, Nvidia is aiming this graphics card at the type of gamers who would buy two or three of these. The similarly priced dual-GPU GeForce GTX 690 performs about the same as the GTX Titan, but is louder, uses more power and is limited to SLI (so not three or four cards). Since you can combine three or even four Titans, you're able to get more performance out of the Titans than you can with two GTX 690s.

You obviously don't need this much graphical power for most things, but we've come up with a few extreme scenarios where a second, third or even fourth Titan can be very useful. (Read: Crysis 3!)

http://us.hardware.info/reviews/4031...and-frametimes

[zalbard]

Pretty terrible scaling for 3- and 4-Way SLI of Titans. And these are all commonly reviewed titles.
Interesting energy consumption numbers, though. But guess it's due to the fact that Metro cannot fully load all the cards.

[kromosto]

sorry to ask but what is that frametimes in the test? for example it writes 7.2 ms what should i understand from this? if it is the average frame render times of the test then it shows us nothing as all 10ms and one 5 one 15ms frames gives us the same result and shows us nothing.

[cx-ray]

According to them it's the time it requires for 99% of the frames to render.

[vitchilo]

Now... what about doing a review on how much 4 GTX Titan can do WCG point in a day??

[naokaji]

Pretty terrible scaling for 3- and 4-Way SLI of Titans. And these are all commonly reviewed titles.

Well, 4 way sli is apparently not even officially supported with titan as the review mentions, plus its not like the competitors solution would do better: http://www.guru3d.com/articles_pages...review,26.html (look at the text below the charts for 4 way xfire, negative scaling, crashes and so on).

Really seems like anything beyond 3 cards from either the green or red camp is more hassle than its worth.

[demonkevy666]

Well, 4 way sli is apparently not even officially supported with titan as the review mentions, plus its not like the competitors solution would do better: http://www.guru3d.com/articles_pages...review,26.html (look at the text below the charts for 4 way xfire, negative scaling, crashes and so on).

Really seems like anything beyond 3 cards from either the green or red camp is more hassle than its worth.

the more powerful the cards get the worse the scaling is getting? :/

sometimes I wish people would go back and test scaling of lesser cards in 4 way, but now that where limited to high end to it I guess it's pointless now.

I've loved see old benches of 4 way 5670 and 5770, silly yup, informative possibly.

seems like lots of games are gpu bound still.

[Elmy]

http://us.hardware.info/reviews/4031...and-frametimes

Best review I have seen so far.

[kromosto]

i still don't get what we should understand from the framtimes.

[Final8ty]

I would hold judgement and see how the drivers progress.

[tajoh111]

Man, so many website are doing tri-SLI reviews. These cards can't be that limited.

[mutantmagnet]

Best review I have seen so far.

Anand's was the best but far from perfect. I really like how this website presents their data.

[paulbagz]

Man, so many website are doing tri-SLI reviews. These cards can't be that limited.

They're not.

nVidia said there should be plenty of stock for these, probably just suppliers having trouble getting them in the initial craze.

-PB

[cadaveca]

i still don't get what we should understand from the framtimes.

1000 ms in a second.

7ms = max FPS of 142.87.

The distance between frametime FPS and actual FPS shows internal GPU efficiency, or signs of other issues and a mis-balanced system.


Also, we must look at differences in frametime when adding multiple GPUs..if frametime increases, then adding GPUs is rather pointless(can add to input latency, ruins internal efficiency). We should look for less frametimes with multiGPU, like by 1/2(1+1 = 2, so each GPU should do half the work, and with proper sync, would halve frametimes, with slave GPU putting it's frames in between the frames of the first, I hope that makes sense to you), and anything less than that indicates an issue.


The root reason for looking at things like this is that FPS doesn't tell the full story. IF you get twice the FPS, but the same frame time, then the end experience while gaming is the same, assuming that you already have met refresh budgets(>60FPS).

There is a big disconnect from getting more FPS, and getting a smoother experience. OEMs currently seem focused on FPS, and scores, rather than actual experience, and AMD has publically admitted they never considered that smoothness might be more important than performance, but I don't blame them, since I am sure they got that info from running surveys, but did not explain what they were really asking, honestly. It's very hard to program drivers for good multi GPU, and if everyone asks just for FPS ,that gives them the easy way out....and all the problems with Crossfire you have today.

There's no need in my mind to discover why AMD fired driver staff when this is the current situation.

[kromosto]

thank you actually i had to be more clear i think. i know about what you wrote but what i didn't get is what should i suppose to understand from 7ms that is different from fps numbers. if it is average frame time then 1000/7ms must give us exactly the average fps, if it is the minimum frame time then it gives us the max fps and vice verse we should have the min fps. there is a 99 percentile thing that they said they don't put worst %1 frame times in the test but this even means nothing more then max min and average fps numbers just a little better numbers thats all.

also lets look at the tests ac3 1920*1080 very high titan fps = 108.7 1000/108.7 = 9,199ms but in test it says 10.6ms in 99p. actually if they ignore worst %1 frames then 99p value must be better then other but it is just the opposite. so i really can't get how they calculate that 99p number.

if they are trying to give us the smoothness of the game then they should give us the delta ms of rendering times in a test.

[El Mano]

The 99% thing is rubbish. What really matters is that 1% that ruins the experience. It's that slow 1% that causes the stuttering.
The 99% is almost the same as the average fps. The slow 1% is what matters. If that 1% is much higher than the 99% then there will be stuttering, no matter how low is the 99% average.

[dasa]

interesting to see a unmoded game taking a huge hit with a 2g gpu although it does take 5760x1080 4xaa max detail to happen
gtx690 sli goes from matching titan tri sli at 92fps medium detail no aa down to 5fps while titan is still pulling 42fps

[CryptiK]

^And that is exactly why some people should buy a Titan vs a 690

[kromosto]

ok i think i figured it out . this frametime shows the worst frametime after ignoring the worst %1. their logic is if worst frametime is better then something then it is unnoticeable so no problem. but actually ignoring worst %1 ruins this test. also constant 40fps with same frametime (25ms) for every frame can produce a better game play then fluctuating frametimes which has a worst frametime of 20ms (which if constant gives 50fps but as this is the worst frametime then average fps is >= 50 fps). giving the worst frametime after ignoring worst %1 times is useless in my book.

[STEvil]

Just look at the crazy frametimes 4-way titan gives..

[zalbard]

The 99% thing is rubbish. What really matters is that 1% that ruins the experience.

"99 percentile frame time" is a latency testing replacement for average FPS.
There is a reason Techreport also includes "number of frames over X ms" graphs.
You cannot stick to a single figure here, the issue is far too complex.

The 99% is almost the same as the average fps.

LOL, no. It's not.





They both matter, really.

[El Mano]

You cannot stick to a single figure here, the issue is far too complex.
They both matter, really.

Yes, and I agree, but review sites are doing it wrong by removing that 1% out of the equation. That 1% is not an statistical error, it actually happens and it ruins your gaming experience. They should focus on that 1%, and perhaps even on the 5% slower frames.
Removing that 1% is just like saying "this is how the game would feel if there was no stuttering". It's unreal, fictitious. It is not what you are going to experience when running the game at home .

[STEvil]

"99 percentile frame time" is a latency testing replacement for average FPS.
There is a reason Techreport also includes "number of frames over X ms" graphs.
You cannot stick to a single figure here, the issue is far too complex.

Its not a replacement and the usage of it as such is incorrect. Frametime measurements are to show deviation in framerate which indicates poor perceived smoothness to the user. By cutting out the 1% you are negating irregularities in performance and thus negating the exact point of the test. If you want to use a 99th percentile of something you should use it on high/low/avg FPS. Also most (in fact all in professional GPU reviews for gamers so far) frametime benchmarks lack the inclusion of a high/low peak latency graph to tell how often per X frames the peak was reached.

LOL, no. It's not.





They both matter, really.

Yes, they do matter, when done right. The average fps and 99th percent frametime numbers line up in many situations so far though and lose their effectiveness in use because of previously mentioned information. That said the frametime issues are not always due to just the GPU and/or its drivers. Many times the game engine or the PC configuration itself is at fault.

[rsnubje]

For those who don't understand how to interpret the frametimes, please read this article.

[kromosto]

For those who don't understand how to interpret the frametimes, please read this article.

i read it but the way they are giving us the frametime information is wrong.