Platform Comparison and Closing Thoughts
To give us a different view of all this data, I decided to make another set of graphs that only compared the CrossFire results of both 7850K and 3960X platforms, removing the single GPU performance for clarity.
There are some issues to work out with Battlefield 4 frame pacing in Mantle. At both resolutions, the system with the A10-7850K is actually a little bit faster than with the Core i7-3960X (orange line vs the green line). That is definitely NOT the case with DirectX. At 2560×1440 everything is much more tightly packed together as you would expect, where the GPU is a more dominant bottleneck.
Closing Thoughts
I think its clear that with Battlefield 4 there is still work to be done with the Mantle multi-GPU implementation. Both AMD and DICE are forthcoming about this, but I thought it was important to run a set of tests today to have a baseline to measure improvements going forward. There are great signs of what direct access APIs can offer for multiple GPUs in a system – the reliability of frame times with the Core i7-3960X is astounding and bests anything we have seen on DirectX from either AMD or NVIDIA. But this appears to be at the expense of raw performance. What we really want is a balance between the two – something like what NVIDIA has proven it can do with SLI.
With Mantle and the upcoming DirectX 12 API taking advantage of the low level access to the GPU, developers are going to be more responsible for multi-GPU implementations than ever before. What we are seeing here today from a world-class developer like DICE could be indicative of what we will see moving forward until something is standardized or ironed out. BF4's Mantle multi-GPU implementation is kind of messy still, though we can clearly see the progress from February and the potential moving forward.
I also want to heap praise upon DICE and Johan Andersson for taking the honest route and implementing an FCAT-style overlay for the press to utilize to measure performance and frame pacing. They didn't have to do this, and doing so allows us to see some warts (on BF4 and AMD Mantle) that would have remained hidden for a while longer. But they understand, as we do, that properly discussing and showcasing problems pushes fixes forward with more speed. As a community, we are lucky to have a developer willing to go out on a limb – for Mantle, with G-Sync, with frame pacing, etc., regardless of the allegiances to other parties.
I wouldn’t say that at
I wouldn’t say that at 2560X1440 Mantle results (3960X) “scale up within range of the DirectX 11 CrossFire scaling”. It is more like they stay flat and all other results just go down. The fps are almost identical with the results at 1080p.
And I really don’t see why the conclusion that the results under Mantle are “less than desirable”. The whole thing about frame rating, the reason, I guess, why all this fuss was ever started, was because smoothness is more desirable than frames per second. Has something change here suddenly? I believe not, but I see uncertainty. Maybe now that things are getting better we are starting to ask for more?
Other than that thanks again for all these articles(the fuss lol) that benefit all of us.
Smoothness is definitely the
Smoothness is definitely the key, but you CAN have smooth frame rates and increased frame rates at the same time. We've seen it done correctly in DX11 for quite some time.
Well, this all explains that
Well, this all explains that while I get much higher FPS in Thief while using mantle, it just doesn’t ‘feel’ very smooth.
Thanks for the article.
Was
Thanks for the article.
Was this testing done after the March 31, 2014 patch? I only ask because it lists specific Mantle frame pacing updates.
Patch notes here:
http://battlelog.battlefield.com/bf4/forum/threadview/2955065227487387474
Yep, we ran all these tests
Yep, we ran all these tests yesterday!
The Multiplayer results look
The Multiplayer results look weird; if I’m reading this right, at 1080p, the Crossfired DX11 solution is the slowest. Looks slightly better at 1440p, but that’s likely only because the single cards are bottlenecked. This looks like a driver issue to me, as there’s no reason CFX should be doing that badly in DX11.
I meant AMD results; not
I meant AMD results; not MP.
Also worth noting AMD’s poor D3d performance in Thief, especially when compared to Mantle.
I don’t want to say AMD is intentionally reducing the performance of their D3D driver, but some of these results sure make it LOOK that way.
It does look that way,
It does look that way, indeed.
Yea I noticed that in theif
Yea I noticed that in theif as well, AMD blog that had graph’s that shows difference on their card from DX11 to mantle yes helped they cards but how a 290x only avg’s 55fps I think was what they said. My 780 (no not ti card)was hitting almost 70fps avg using same settings. If it ends up coming to Light AMD was dumbing down DX11 performance on games with mantle to make mantle look better, could be nasty PR nightmare. Likely since game is a mantle game they don’t want to focus on optimizing the game for dx when it competes with their own proprietary api and make results look less compelling.
other difference is I have a 4770k not 4670k, but results should be to much different.
http://community.amd.com/servlet/JiveServlet/showImage/38-2668-2822/290X_4670K.png
Great catch.
Great catch.
It’s not true though, 290x
It’s not true though, 290x easily averages over 60 in DX.
Perhaps you’re quoting data from the older drivers, or kids with incredibly broken ‘enthusiast’ setups.
well it seems like a
well it seems like a pointless wast of resources, to spend time optimizing for dx, while they can focus on mantle, if it was me i would do the same.
Why fix SLI frame pacing when
Why fix SLI frame pacing when you can just launch G-Sync.
Nvidia never had SLI frame
Nvidia never had SLI frame pacing problems, Nvidia fixed that problem back at start with Hardware on the cards to do it. frame pacing was all AMD’s problem.
Of course they have frame
Of course they have frame pacing problems! Look at the SLI frametimes! A complete mess!
Mantle is an almost perfectly flat line, NVidia cannot achieve that with dx.
Why fix SLI frame pacing when
Why fix SLI frame pacing when you can just launch G-Sync.
Because G-Sync is not free of
Because G-Sync is not free of charge. It is expensive and only in few(are there more than one?) monitors.
Because Mantle will be
Because Mantle will be obsolete since DX12 does the same thing, runs on more platforms (Windows + Windows Phone + XB1), and runs on more cards (NVIDIA + AMD + Qualcomm). And then when NVIDIA’s cards are 20+ FPS faster, I’m sure AMD will be sorry they didn’t fix their D3D driver up.
Except, Mantle will run on
Except, Mantle will run on every console, the steam box and anything that can run a Linux distribution. As well as Android.
nVidia had frame pacing
nVidia had frame pacing sorted way back in 2008. AMD in 2014 have finally got it sorted for their user base. G-Sync is completely different to Mantle and can’t be compared.
Not only sorted out, but
Not only sorted out, but without any extra APIs. Shows who can develop drivers and who can’t and has to forced developers (often crappy themselves) to do job of driver vendor…
NVidia did not have frame
NVidia did not have frame pacing “sorted” in 2008. They have been aware of the issues and working to solve them with a slow, steady improvement, however it’s not completely perfect even today.
I watched a really long video with Tom Peterson at PCPER (90 minutes?) and frame pacing was discussed quite a bit, including SLI frame pacing.
In particular, Frame Rate Metering was discussed and he said they will continue to improve it every generation in both hardware and software.
(I also wonder if GTX750Ti cards don’t have SLI in part because new Frame Rate Metering might not be finalized for Maxwell cards.)
Why would the author
Why would the author benchmark on older 14.2 Mantle drivers? There is clearly something wrong here with crossfire scaling. I get a massive boost only 7970GHZ Crosfire and 3930K.
Actually, that was a typo. We
Actually, that was a typo. We used Catalyst 14.4 beta drivers we got in house on the 31st.
Thanks for the clarification.
Thanks for the clarification. 14.4 is not released yet. I thought you would make it clear if you were using a pre-release driver. Also those kind of things customarily require an NDA. Amazed AMD allowed you to post benchmarks.
Ryan, have you messed with
Ryan, have you messed with this setting “RenderDevice.FramePacingMethod” at all?
The default setting 2 is cutting my fps in half on my 7990 its crazy, what setting did you use for the benchmarks here?
“Smoothness is definitely the
“Smoothness is definitely the key, but you CAN have smooth frame rates and increased frame rates at the same time. We’ve seen it done correctly in DX11 for quite some time”
Cool…I`m still rockin` a AMD 5850/C2Q and it works fine.
It was one of the first DX11 cards got it to play Dirt 2 !
Still is fun to play.
Using 14.2 beta 1.3 with
Using 14.2 beta 1.3 with crossfire 290’s on bf4 and mantle is the smoothest gaming i’ve seen in years and this is the first time in the last 10 years or so that i’ve owned non nvidia gpu’s. Seems like a lot of the testing i read on different sites somehow never relates to my experiences and i’m having trouble understanding why this is. The dual gpu scaling is very impressive with upto 100% scaling and i can honestly say, i could not be more pleased with the 290’s.