Further Anti-Aliasing Investigations
When it comes to anti-aliasing in AC, the heart of the matter is the differences in graphics architectures. ATI's current GPUs do not have dedicated anti-aliasing hardware. Instead, they rely on the pixel shaders to achieve the same result. Depending on the type of graphics work being done, this approach can have a negative impact on performance. ATI seems to have gotten around this via some of the extensions to DirectX 10.1. NVIDIA graphics chips do include anti-aliasing hardware, but that hardware cannot be properly utilized in some situations. Specifically, there are post-rendering effects that interfere with the use of anti-aliasing hardware -- that's why some games don't support anti-aliasing at all.
If you haven't figured it out already, there is a universal "solution" for applying anti-aliasing effects that doesn't depend on the use (or non-use) of other shader effects. The solution is to use pixel shaders. ATI has apparently decided that's the most practical solution as we move forward and SM 3.0/SM4.0 usage increases, but it does require work on the part of software developers. Another downside to this approach is that it requires more work from the pixel shader hardware, which can result in lower performance. However, it seems the only other option is to omit anti-aliasing in certain types of graphics rendering.
With all that out of the way, let's look at the performance of AC with and without anti-aliasing at several different resolutions. Unfortunately, as previously mentioned we are unable to test anti-aliasing at higher resolutions because the option is disabled inside the game.
We can see the result of using pixel shaders to do anti-aliasing in the resulting performance drop. What's noteworthy is that the drop isn't nearly as bad on ATI hardware running AC version 1.00. In other words, the 1.02 patch levels the playing field and forces ATI and NVIDIA to both use an extra rendering pass in order to do anti-aliasing. That probably sounds fair if you're NVIDIA -- or you own NVIDIA hardware -- but ATI users should be rightly upset.
Something else that's interesting to see is how performance is definitely CPU limited even with a quad-core 3.0GHz Intel chip. We will look at that next, but right now we are more interested in the steady drop caused by anti-aliasing. If the CPU is the bottleneck, putting more of a load on the GPU should not impact performance much. For whatever reason, it seems that the extra rendering pass required for anti-aliasing causes a steady drop in performance even when we're CPU limited. Have we mentioned yet that you really need some beefy hardware in order to play AC at higher detail settings?
Facebook
Twitter
RSS
32 Comments
View All Comments
bill3 - Monday, June 2, 2008 - link
Actually it's terrible, I cant read the graphs AT ALL.seriously my eyes just glazed over those terrible charts..completely unreadable. I still, have no idea what I'm looking at. Is ATI supposed to be faster in this game? Why did they test with version 1.00 on ATI and 1.2 on Nvidia? I dont know because the graphs are totally useless.
Nihility - Monday, June 2, 2008 - link
I second that. The graphs are terrible. Maybe bar graphs would have been better?Sometimes when you're the one making the graph it's hard to imagine what other people are seeing when they look at them. I suggest having another pair of eyes check the graphs out for readability.
Besides that, I loved the review. Especially the performance part and the 10.1 controversy.
JarredWalton - Tuesday, June 3, 2008 - link
Charts are colored with similar colors used either for ATI vs. NVIDIA, 1.00 vs. 1.02, or dual-GPU vs. single-GPU. I could have generated four times as many graphs to show the same data, but I figure most people are capable of reading the labels on a chart and figuring out what they mean. Here's a hint: when you can't see the difference between two lines because they overlap, it's a tie.If you want to give specific examples and recommendations on what would look better and still convey the same amount of information, I'm all ears. However, simply stating that "the graphs are terrible" does little to help. Tell me what graph specifically is terrible, and tell me why it's terrible.
As an example of why I used these graphs, page 9 has two charts showing 40 total data points. You can get a clear idea of how performance scales with single or dual GPUs at the various detail settings looking at a single chart. Green is NVIDIA, Red is ATI. That makes a lot of sense to me. Creating ten different bar charts with four lines in each to show the same data makes it more difficult to compare how Medium graphics compares to High graphics performance, and it takes up five times as much space to tell the same "story".
Page 6 is the same thing, but with green used for dual-GPUs (light and dark for 1.00 and 1.02) and red for single GPUs. 24 data points in two charts instead of using six charts. Having established that 1.00 doesn't perform any different than 1.02 on NVIDIA hardware, I skipped the 1.00 NVIDIA numbers to make those charts easier to read on page 7. Then I put in the four standard test system (0xAA and 4xAA, ATI and NVIDIA) on 1.02, with 1.00 4xAA ATI in blue as a reference.
Lastly, on page 8 I have two clock speeds on NVIDIA, three on ATI, with different base colors for single and dual GPUs. ATI and NVIDIA are in separate charts, and brighter colors are for a higher overclock.
There's method to my graphing madness. Are the charts immediately clear to a casual glance? No, but then that's really difficult to do while still conveying all of the information. I spent a lot of time trying to make comprehensible charts, and settled on these as the best option I could come up with. Again, if they're so bad, it must be easy to generate something clearly better - have at it, and I'll be happy to use any sensible suggestions. However, if the only complaint is that you actually have to look at the charts and think for a minute before you understand, I'm not likely to be very sympathetic. I think our readers are smart enough to digest these graphs.
mpjesse - Monday, June 2, 2008 - link
While I appreciate the detailed review, isn't it a little irrelevant now? I mean, the game's been out for nearly 2 months now and it's been reviewed everywhere. The only thing new about this review are the performance benchmarks, in which case I would have have made the review solely about performance instead of gameplay.Just my 2 cents.
ImmortalZ - Monday, June 2, 2008 - link
Its sad that the companies with money always manage to suppress innovation.I hope this article by AT will raise some ruckus in the collective Interwebs and cause something. But I doubt it.
ViRGE - Monday, June 2, 2008 - link
For what it's worth, another forum I read had some screenshots comparing DX10 and DX10.1. The problems the poster had managed to find involved trees; there was some kind of post-processing rendering going on with trees that wasn't occurring with DX10.1, which made them look weird.Not fixing 10.1 may be an NVIDIA thing, but there was definitely a problem with it as-is.
tuteja1986 - Monday, June 2, 2008 - link
Well why where the hell is nvidia dx10.1 support if dx10.1 actually brings some kind of performance improvement in AA.Why aren't GT200 series have DX10.1 ?
I thought PC gaming was all about being the cutting edge on all technology front...
Anyways , this is not the 1st time Ubisoft or Nvidia have done this.
wyemarn - Monday, June 2, 2008 - link
Maybe because Nvidia GPUs cant support AA through shaders. So no use supporting dx 10.1. ATI GPUs have 320 stream processors so it can utilize for shaders and etc. Nvidia cards have less SPs but more ROPs, TMUs which translates to more brute power if games dont use shaders or SPs much. Technology wise, I think ATI is ahead but NVIDIA GPUs have game developer support and more raw horsepower so performance wise NVIDIA is ahead and I think this trend will continue with GTX200 series. I choosed G92 over RV670 because the raw performance is much better even though on paper HD 3800 series look great.SteelSix - Monday, June 2, 2008 - link
Worthy of a thread in Video. I just started one..Gannon - Monday, June 2, 2008 - link
The original halo had performance issues but they weren't alarming, halo was actually not too bad port compared to many other console to PC disasters. Halo 1 got 'better with hardware' advancing. Halo 2 on the other hand is just all around atrocious. Halo 2 was just not a very well made game, period, despite the addition of cutscenes, etc. Halo 1 had a much better feel and better vehicle design IMHO, I hated how the warthog looked in Halo 2, it annoyed me to no end.