Geforce GTX280 & Radeon HD4870X2 AA Scaling with XP & Vista

Videocards/VGA Reviews by jmke @ 2008-09-04

In this in-depth article we take a look at the performance of the NVIDIA Geforce GTX 280 and ATI Radeon HD 4870 X2 when anti-aliasing is enabled. We test 8 different games at several different AA levels under Windows XP as well as under Windows Vista. How does performance scale when you go from XP to Vista, how much impact does enabling AA have? Read on to find out!

  • prev
  • next

Crysis

Crysis

Crytek became famous with their Far Cry first person shooter game, not only for the open-ended gameplay but also because of the stellar system requirements to be able to play the game at high detail. Crysis is their second game and doesn’t disappoint in either gameplay or system requirements.

Crysis offers several methods to test performance with their game, they include two batch files, one geared toward CPU testing, the other toward GPU testing. These two methods provide very repeatable results but unfortunately don’t reflect real gameplay performance, and only give you an indication of how the game will run.

We used this Crysis benchmark tool which enables you to define a custom time demo; using the build-in “Assault” run-through we measured performance of both cards. We briefly tested the 64-bit executable included with Crysis when installed on a 64-bit OS, but performance was actually lower with no IQ benefit, so we stuck to the 32-bit .exe for all our tests.

First with High Quality preset:

Madshrimps (c)


As the first reviews of the HD 4870 X2 hit the web we read through most of them and found the performance numbers of Crysis to differ quite a bit between reviews. We did a small report on this here. Now we have tested it for ourselves, and as you can see from the results the choice of OS and Render path is crucial to the outcome of the benchmark.

With the Geforce GTX 280 going from XP to Vista doesn’t drop performance on bit, but enabling the DX10 render path is fatal as you can see, especially with AA enabled the FPS drops below playable levels.

The Radeon HD 4870 X2 fares better, going from XP to Vista with DX9 actually boosts performance by ~30%! Switching to the DX10 render path has minimal impact without AA, and no impact with AA enabled.

Next up, Very High Quality and 2gb vs 4gb testing:

Madshrimps (c)


Neither of these cards is fast enough to run Crysis with Very High preset at 1920x1200, below 30fps numbers everywhere, the HD 4870 X2 is almost twice as fast when AA is enabled, but still too slow. Adding more system memory (2gb -> 4gb) does nothing for the GTX 280, but with the Radeon we did see small fps increase.

For a detail view of the results, with AA scaling and XP -> Vista Scaling see this table




You must also have noticed the stellar performance when it comes down the AA scaling with Crysis; according to the results could enable 4xAA or 8xAA and not notice a drop in performance… to see whether the benchmark was telling us the truth we loaded our own manual FRAPS run-through of the same Assault map; do note that a slightly different path was taken. These tests were done with the HD 4870 X2 and system 2gb ram:

Madshrimps (c)


At the high quality settings the findings are pretty much on par, the difference between 4xAA and 8xAA is indeed negligible. But unlike the results with Very High quality would like you to believe, there is a larger performance drop going from 4xAA to 8xAA here. We’ve read some claims of people playing Crysis at 1920x1200 Very High detail with 16xAA enabled at 30fps + with the Radeon card… but after several weeks of testing we just can’t seem to duplicate this, as soon as AA is enabled performance drops below 30fps when Very High preset is used.
  • prev
  • next
Comment from Kaotik @ 2008/09/06
You forgot to mention that
...8xAA on ATI should be compared to 8xQAA on nV, not the 8xAA which is 4xMSAA based CSAA mode
...16xAA on ATI effectively turn the card into single chip card which can do 16xMSAA, since both chips render the same frame with different AA patterns
...16xAA on nV is 4xMSAA based CSAA mode and 16xQAA on nV is 8xMSAA based CSAA mode

So 16x and 8x comparisons in your graphs are far from being 'fair' or 'apples-to-apples', the 8xAA should have ATI 8xAA vs nV 8xQAA (8xQAA = 8xMSAA) and 16xAA shouldn't even exist since the GTX280 can't do 16xMSAA which is (practicly) what the HD4870X2 is doing by blending the same frame rendered twice with different AA patterns.
Comment from jmke @ 2008/09/07
Thank you for your input. much appreciated; the 16xAA on the ATI does indeed split performance in two, really heavy hit on performance, but ultimate image quality.

While the performance on the two cards at different AA levels can be compared head to the head in the charts, their scaling going from XP to Vista was definitely my main focus.

I don't know if you were the same person who email me about this same issue, but I agree with you that NVIDIA and ATI each have their own approach to the AA levels defined in their control panel, making it less than straight-forward.

on the topic of AA levels, I was hard pressed to find major improvement on NVIDIA going from 4xAA to 16xQAA (through NV control panel); same goes for ATI where 4xAA or 8xAA did little to improve image quality visually, meaning that the to be able to spot the difference you'll have to do screenshot comparisons of static screens and use a zoom-loop in order to find them.

on the upside, we finally almost get free 4xAA under Vista with ATI which will be a key feature for full DX11 compliance if I remember correctly.

Again thank you for your input and the educative message!


-----

I received the following mail regarding Crysis performance with the HD 4870 X2 (and Crossfire)

Quote:
Originally Posted by Felipe by email
I saw you article about scaling with XP & Vista comparing tow of the fastest cards in the market and I just want to tell you, thar, there is a list of Cvar in Crysis that increase FPS when using with a CF setup, this worked with 3870X2 and 4870X2, running 32bits executable.

I can say that Crysis some how is bottlenecking CF systems.

One of the Cvars is R_TEXTURESTREAMING.
Run #1- DX9 1280x1024 AA=No AA, 32 bit test, Quality: High ~~ Overall Average FPS: 46,97
Run #2- DX9 1280x1024 AA=4x, 32 bit test, Quality: High ~~ Overall Average FPS: 37,12
Run #3- DX9 1280x1024 AA=8x, 32 bit test, Quality: High ~~ Overall Average FPS: 36,33
TEXTURE STREAMING = 0
Run #1- DX9 1280x1024 AA=No AA, 32 bit test, Quality: High ~~ Overall Average FPS: 56,11
Run #2- DX9 1280x1024 AA=4x, 32 bit test, Quality: High ~~ Overall Average FPS: 42,41
Run #3- DX9 1280x1024 AA=8x, 32 bit test, Quality: High ~~ Overall Average FPS: 43,00
And using this combination here (based on my system)
sys_budget_numdrawcalls = 6000
sys_budget_videomem = 512
sys_budget_sysmem = 6144
sys_budget_frametime = 80
r_TexturesStreaming = 0
Using Only texturestreaming
Run #1- DX9 1680x1050 AA=No AA, 32 bit test, Quality: High ~~ Overall Average FPS: 4635
Using Texturestreamin and Budget config
Run #1- DX9 1680x1050 AA=No AA, 32 bit test, Quality: High ~~ Overall Average FPS: 4869

This have the same effect on 4870X2 and 4870 series, for some reason it doesnt help Nvidia cards.
Comment from Kaotik @ 2008/09/07
Nah, I'm not the same person, it was another Beyond3D user as far as I know who mailed you.

I just thought that those differences should be mentioned even though the comparison between Vista & XP was the main concern of the great article.

Regarding the 'free 4xAA' on HD4870X2 you mentioned, I think you're mixing it with the DX10.1 requirement of 4xAA, which has nothing to do with performance, just that you have to have the support for it (as far as I know, DX10.1 was the first DX ever to actually require support for specific AA mode)
Comment from jmke @ 2008/09/07
I've added an addendum to the article regarding the difference in AA levels between ATI and NVIDIA

I thought DX10.1 (and DX11) allowed 4xAA without a performance hit, much like Assassin's Creed on the ATI cards before they released the patch
Comment from Kaotik @ 2008/09/07
Quote:
Originally Posted by jmke View Post
I've added an addendum to the article regarding the difference in AA levels between ATI and NVIDIA

I thought DX10.1 (and DX11) allowed 4xAA without a performance hit, much like Assassin's Creed on the ATI cards before they released the patch
Nope, there's no such thing as "free AA" really even though the performance drops are relatively small today. DX10.1 (or 11) changes nothing on these regards, 10.1 just introduced the requirement to have support for 4xMSAA to be 10.1 compliant, and allows access to the samples "better" (no idea how to better describe it really)
In case of Assassin's Creed, they took advantage of DX10.1 how they can access the AA samples later in the rendering too, with DX10 they had to do two rendering passes for some effects, one with and one without AA, while with DX10.1 they could just re-use the buffers from AA rendering pass, completely dropping out one rendering pass.

 

reply