Review Index:
Feedback

Frame Rating: GTX 660 vs HD 7870, plus HD 7790, HD 7850, GTX 650 Ti BOOST

Author: Ryan Shrout
Manufacturer: PC Perspective

Battlefield 3 - GTX 660 vs HD 7870

Battlefield 3 (DirectX 11)


 

Battlefield 3™ leaps ahead of its time with the power of Frostbite 2, DICE's new cutting-edge game engine. This state-of-the-art technology is the foundation on which Battlefield 3 is built, delivering enhanced visual quality, a grand sense of scale, massive destruction, dynamic audio and character animation utilizing ANT technology as seen in the latest EA SPORTS™ games.

Frostbite 2 now enables deferred shading, dynamic global illumination and new streaming architecture. Sounds like tech talk? Play the game and experience the difference!

Our Settings for Battlefield 3

Here is our testing run through the game, for your reference.

View Full Size

View Full Size

Right away with the Radeon HD 7870 we see a difference between the FRAPS measured frame rates and the observed frame rates of our Frame Rating capture based analysis system.  In many areas we found the performance of the CrossFire configuration to be no faster than the single GPU, when we see large amount of alternating full frames and runt frames. 

Worth noting of course though is that the single GPU results are showing the HD 7870 as the superior card compared 1 on 1 with the GTX 660.

View Full Size

Our frame time diagram shows where those drops in observed frame rates happen and WHY we measure them that way.  The GeForce GTX 660s in SLI though result in a pretty steady frame time sequence with very little variance.

View Full Size

This graphs shows the percentile distribution of frame rates, effectively your minimum frame rate level at any given percentile.  You can see that the "average" frame rate of the SLI configuration jumps from 59 FPS to 110 FPS or so, while the CrossFire only increases from 61 FPS to 80 FPS. 

View Full Size

Our data shows a high level of frame variance on the CrossFire result as you would expect based on our frame time plot above. 

 

View Full Size

View Full Size

The same pattern is seen at 2560x1440 as we witnessed above at 1920x1080 - the CrossFire result sees significant runt frame counts and because of that the "observed" frame rate drops noticeably.  Also, the GTX 660 is running consistently slower than the HD 7870 in a single GPU scenario.

View Full Size

The frame time graph looks even worse at 25x14 than it did above with more frequent episodes of alternating runt and long frames displayed to the gamer.  The SLI results do show a bit more variance as well though nothing close the ballpark of the CrossFire numbers.

View Full Size

Comparing the orange (CF) an the black lines of the HD 7870 there is very little difference between the observed performance based on FPS percentile data.  In fact, towards the end we see the CrossFire results dropping under that of a single card.  The GTX 660s running in SLI though are much better with the average FPS scaling from 36 FPS to 68 FPS, an increase of 88%.

View Full Size

Clearly the frame time variance and potential stutter seen in the Radeon HD 7870s in CrossFire are a problem, shown here in another light.  The SLI configuration does see more variance than the single GPUs but it doesn't really get bad until we hit the 95th percentile.

 

View Full Size

View Full Size

At 5760x1080 running in either AMD Eyefinity or NVIDIA Surround, the HD 7870s in CrossFire exhibit more problems, this time in the form of dropped frames.  Once we take those out of the calculations we see performance at or below the results from the single HD 7870.  NVIDIA's GTX 660s in SLI though scale consistently as you would expect.

For single card results though, the GTX 660 is definitely falling behind the HD 7870.

View Full Size

That blob or red shows you the dropped frames that are missing from the colored overlay that we expected to see.  One of the Radeon HD 7870 GPUs is basically doing work that is never seen by the user...

View Full Size

The CrossFire results look really bad in our frame time plot but even the GTX 660s in SLI are showing some higher than expected frame time variance.  Compare that to the slower, yet more consistent frame times on the single HD 7870 and GTX 660 and you can see why multi-GPU simply produces more complicated issues for developers and GPU vendors. 

View Full Size

Our frame rate graph definitely indicates that SLI is scaling with an average frame rate that increases from 22 FPS to 37 FPS but that SLI does tail off at the end.  In the worst cases here (99th percentile) we only see scaling jump from 16 FPS to 20 FPS.  CrossFire never is able to separate from the single HD 7870 results.

View Full Size

Because we remove the dropped frames all together from the CrossFire data, it appears that the frame variance levels of CrossFire HD 7870s is LOWER than the GTX 660s in SLI - and that is true in the purest sense.  Frames that never show up on the screen can't "vary" from the running average of previous 20 frames!  

April 5, 2013 | 07:29 PM - Posted by Torrijos (not verified)

Great work! I guess you're as happy as your readers to get to the end of this generation cards.

It would be nice, in order to facilitate readers choice, to have a summary page with all cards presents (maybe with dynamic selection).

What I image would be best is the graphs, where the reader selects a Game, a Screen Resolution, and gets all the cards on the same graphs (with tick-boxes to removes the cards he isn't interested in).

Graph= ƒ(Game, Screen resolution, GPU selection)

That way a reader looking for the best performance for his screen and favourite game would have all the information needed on a single graph. (you could be motivated to create another graph à-la TechReport taking into account the price)

Anyway Thanks again for the good work!

April 6, 2013 | 02:09 AM - Posted by Luciano (not verified)

thumbs up for this suggestion!
thats what here in brazil they call "info-graphics" (redundant i know by you get the idea).

April 5, 2013 | 08:28 PM - Posted by Cerebralassassin

Great work Ryan!

April 5, 2013 | 08:34 PM - Posted by ThorAxe

Thanks for a another great article. As a few others have pointed out it would be great to see older cards tested such as the 8800GTX, HD 4870 and up. I realize that this would be a huge amount of work but not everyone is a graphics whore like me that upgrades almost every cycle.

April 5, 2013 | 11:21 PM - Posted by Tim Verry

Ah, the good ole' 8800GTX. My little brother is running one of those still and when I finally upgrade he'll get my psuedo-6970 (unlocked xfx 6950). Unfortunately he missed out on my 3850 upgrade when it died on me and would no longer output anything to the display. It plays Kerbal Space Program, Minecraft, and most other games (with lots of stuff turned out of course) so yup, I would like to see Frame Rating through the several generations of graphics cards--even if it's only one card from each NV and AMD's generation and not their whole lineups (heh, we'd never see or hear from Ryan or Ken for 6 months if they had to Frame Rate all those cards :P).

April 6, 2013 | 02:12 AM - Posted by ThorAxe

I still have a HD 3850, 2x 8800 GTX, HD 4870x2, 2x HD 6870s and 2x GTX 570s SLI in various PCs of my wife and kids . My current gaming PC is 2x GTX 680s SLI. I told you I have a sickness. :)

At one point I ran a 4870x2+4870 in Trifire and it didn't seem as bad as the 6870s in Crossfire if I remember correctly, so it would be interesting to see what impact Trifire has on Observed Frame Rates.

April 6, 2013 | 02:25 AM - Posted by Luciano (not verified)

Your sickness gives you the power of intuition:

http://www.tomshardware.com/reviews/radeon-geforce-stutter-crossfire,299...

AMD fans: "see? you just need to buy a third card! im sure is cheaper than 2 nvidias"

April 6, 2013 | 01:32 PM - Posted by bystander (not verified)

Unfortunately only 1 card was tested, so conclusions are hard to be definite. What might have fixed it may be what fixes 2-way Crossfire, a CPU bottleneck. When you add more cards, the GPU puts more stress on the CPU, and if it couldn't keep up, it will bottleneck, which also has shown to fix stuttering.

More testing is needed. Same with V-sync. Only 2 games were used on the v-sync test. It may be that v-sync only helps due to its FPS limiting component, though I have a feeling it would help anyways, but more testing is needed to be sure.

April 6, 2013 | 01:34 PM - Posted by bystander (not verified)

I meant that only 1 game was tested, not one card.

April 6, 2013 | 02:20 AM - Posted by Luciano (not verified)

Maybe its time for you elite of reviews (pcper, techreport, toms, anand) get together to work on the data generation in gpu reviews.
The value of each reviewer is what he points for us to see more closely and his conclusions.
Toms for instance is saying "just turn on vsync and everything would be ok".
PCPER refuted that.
Thats why god gave us multi-tab browsing.

April 6, 2013 | 11:19 AM - Posted by bystander (not verified)

Unfortunately different people have different preferences. You typically find competitive gamers refuse to use v-sync, because it induces latency and you lose the advantage of high FPS.

And curiously enough, v-sync doesn't work with Eyefinity.

April 5, 2013 | 10:32 PM - Posted by bystander (not verified)

@Ryan:
I'm having a hard time finding what you are defining a runt frame. THG has it set to under 21 rows, which they said was the default, are you using the same definition, or something else?

April 7, 2013 | 04:07 PM - Posted by Ryan Shrout

Our versions of the scripts define it based on percentages.  If the frame is 20% or less of the preceding 10 frames average, then we consider it a "runt".

April 7, 2013 | 05:11 PM - Posted by bystander (not verified)

Cool, that is actually a cool method. The higher the FPS, the smaller the frames will be shown, so having it scale based on the average is good.

April 11, 2013 | 04:06 PM - Posted by Anonymous (not verified)

With 1080 horizontal "scan lines" on the average screen, calling 21 scan lines a frame is a joke really.

25% of one full screen would be 270 scan lines. Calling a quarter of a screen a "frame" is stretching it quite a lot.

Maybe one fifth the screen is a reasonable cutoff point, or one fourth. Anything less is a joke really, as once 5 "frames" are "sharing" one screen, we see slivered runts in all the examples given.

So I expect some reason for the answer, but I suspect we'll get ridiculous things like 7 or 8 scan lines out of over 1,000 on the screen "counts" as "one frame" or some tiny portion of the whole screen, far less than 20%, less than 10%, less than 5%, as 7 or 8 scan lines is less than 1% !

I can do the math, and do the visuals. Good luck arguing me down to 10% of a 1920x1080, or "108 scan lines" counting as one FPS.

April 12, 2013 | 04:42 PM - Posted by bystander (not verified)

I can't tell if you are arguing or agreeing with my thought.

I will add that as your FPS go higher, the smaller the slice of a frame will be on average. At 300 FPS, the average frame size will be about 20% of the screen. This is why I like what they did, as it scales based on the average frame size.

April 6, 2013 | 12:13 AM - Posted by Shambles (not verified)

Amazing timing! Needed exactly this.

April 6, 2013 | 12:31 AM - Posted by lindethier

Good work!

April 6, 2013 | 01:23 AM - Posted by Number1 (not verified)

I might have felt this dragged on a bit too long but I was wrong. It was important to do the tests across the whole line up of current cards. I was mainly concerned that you would end the investigation here, after doing so much work on this, but now that you have stated it is on going I am very impressed. Anyway just thought I'd say good work. Really good work. Very glad to hear that you will be looking at user submitted solutions.

It will be interesting to see if peoples claims about radeonpro etc are true. I have seen people post pretty impressive results in terms of frame latency using these solutions. I have also been promoting this solution going off other peoples reports. Hopefully there aren't too many costs associated with using it (other than setting up individual game profiles, and making sure you do it right..)

I agree with other users in that it would also be great to check say 6950s or 6970s in crossfire. AMD has suggested things like the GCNs memory manager needs work. AMD haven't specifically suggested that this is where crossfire is falling down and I don't think it's related. However, that isn't stopping people around the internet claiming it is the cause though. So it would be interesting to have a single non-GCN card thrown in, even if its only tested in two or three games, just to confirm it is AMD's whole crossfire solution rather than their GCN products and its memory management.

Good work nVidia in the last couple of years for making an SLI solution that works out of the box so well.

Good work to AMD for producing cards that are performing so well in the single card tests. Feeling pretty happy about my 7870 tahiti LE purchase.

April 6, 2013 | 03:33 AM - Posted by alwayssts (not verified)

Believe me when I say I understand that asking for more after all this feels terrible, and am greatful for the work on these articles but I do wish Ryan would test Tahiti LE...as well as show some 'typical' overclocking now that stock comparisons have been made.

It's very simple to argue LE is one of the most relevant cards for the largest market because of it's price/performance and absolute (overclocked) performance at 1080p, even if it's 'limited edition'/highly salvaged (which are usually awkward niche parts) and certainly soon to be replaced by Hainan...which will undoubtedly have a slightly better stock placement relative to 660ti. This is similar to the way '7870' became 7950 in performance through Tahiti LE, 7950B became the old 7970, and 7970GHZ ED was to have a stock config to beat the nvidia competition which often equaled or beat those original configs). Am I the only one telegraphing that coming stock above 660ti (which will drop to 250) with partner overclocks (like 7790 has with 1075/6400mhz) very similar to a stock 670 (but probably priced cheaper? Think 250, 270, 300, 330 (or something like that).

I would say I hope we get a 670/Tahiti LE update article, but I feel 670 is best left for when Hainan launches...and wonder if that is Ryan's thinking as well considering the deliberate choice to leave it out. I do wish both would've been thrown in the 660ti/7950B article, but understand his reasoning. 670 is in it's own price niche, just like 7870XT. I know the later may be a 'tweener, but it really would be interesting to see the results as I think it will show a nice point of reference/cutoff where most people should be looking to get smooth 1080p HQ game-play.

April 6, 2013 | 01:34 PM - Posted by Number1 (not verified)

Personally I think there is enough data to extrapolate how the gtx670 and tahiti LE will perform.

April 6, 2013 | 10:56 AM - Posted by AlienAndy (not verified)

More epic data !

April 6, 2013 | 04:50 PM - Posted by Joshua B. (not verified)

I am interested to see how crossfire/sli are effected by slower processors. I know that there is no end to the types of configurations out there but I wonder if causing a bottleneck elsewhere in the system will cause either adverse or positive results.

I also know that originally the setup used was to show the performance of the cards unrestricted, to which I say is a good thing when it comes to comparison from card to card, but now I find that adding in the cpu (crossfire/sli) may generate different results, unless you feel that the cpu still doesn't cause any "real" problems as far as perceived framerate. ex. the bulldozer review.

April 7, 2013 | 02:48 AM - Posted by Howie Doohan (not verified)

Shocking results from AMD. These results almost completely invalidate any price savings that might be had by going crossfire (eg CF 7970's vs SLI 680's). I'm definitely going green for my next multi card rig and won't be giving AMD a look until they pull their thumb out and fix crossfire.

April 7, 2013 | 09:03 AM - Posted by Humanitarian

Very comprehensive look at frame rating, excellent write up.

Really interested to see what some of software solutions previously suggested about will effect this problem.

April 7, 2013 | 11:34 AM - Posted by keromyaou (not verified)

A great work!!
I would like you to test older hardware (such as HD5870) to see when this crossfire problem occurred.
I found one link which might be interesting for you (http://www.tomshardware.com/reviews/radeon-geforce-stutter-crossfire,299...). Although the test in this link is not sufficient for withdrawing the final conclusions, it suggested that microstuttering was more in HD6870 crossfire than in GTX560 SLI setup and also that triple crossfire setup with HD6870 alleviated microstuttering greatly. Since their methods were not as finesse as yours, they probably were not able to detect some issues (like runts) like you were. But still their data is worth noting, I guess. I think that their data hinted that:

(1) Crossfire issues might be older than HD7000 series.
(2) Triple crossfire setup might be able to alleviate the problem which occurs with dual crossfire setup.

I hope that you have a chance to test triple crossfire setup with your apparatus as well as older gpu cards.

Thank you for your interesting data.

April 7, 2013 | 04:09 PM - Posted by Ryan Shrout

I think we are going to test one or two of the previous generations of cards to see how things scaled throughout.  Right now I am considering the HD 5870 and the GTX 480.

April 7, 2013 | 02:27 PM - Posted by Trey Long (not verified)

Great work. This is exactly the type of investigation that will make ALL companies better and go to a higher level.

April 7, 2013 | 06:27 PM - Posted by Anonymous (not verified)

Great test as always, but I think you should test GTX 580 SLI vs 660 SLI ASAP!
both have around the same performance, but Kepler supposedly deals with frame rate metering with dedicated hardware for that, and the 580 doesn't, both will use the same driver, from the same vendor, so it would be a great test and maybe help to understand how close AMD could get just with software tweaks!?

April 7, 2013 | 10:40 PM - Posted by I don't have a name. (not verified)

That is a huge amount of data presented Ryan. On the whole a killer article, one of the best I've ever read. The idea of testing the older 480 and 5870 should prove very interesting.

Thank you for such an informative, cohesive and thorough and downright geeky article. I loved reading through it. I think I went through the first article about five times just to ensure I had all the concepts fully understood. Awesome. :)

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

By submitting this form, you accept the Mollom privacy policy.