Writer: DGLee


I felt that CFX performance was screwed up when I had just installed Catalyst 11.4 WHQL. For so long, I've been wondering whether new Catalyst is a viral incubus or it's a problem of VGAs themselves. But up to now there's no clear investigation on that issue, I'm to configure this phenomenon with quantatitive measures. Here's the epic result. ;-)
(Note that all games tested below are my own titles - which are purchased properly)

Basically, I ran each game 3 times for every single combination of [CFX mode(3-way, 2-way, single) * Catalyst version(11.2 ~ 11.5)]. For unacceptably low or high results (which seem 'fluctuated' against results from former & latter version of Catalyst), I did again the whole 'ran each game 3 times for......' procedures. So variable control is nearly perfect. Trust me & these results forward. :-)

Underlying spirit of this benchmark is to test games at highest-option & resolution (2560 x 1600). Detailed settings are to be revealed together with graphs.

Test system is as follow:

- CPU: AMD Phenom II X6 1090T @ 4.2GHz / NB 2.8GHz
- M/B: MSI 890FXA-GD70
- RAM: Samsung PC3-10600 4GB x 2 @ 1600MHz CL7
- VGA: AMD Radeon HD 6970 2GB x 3
- Storage: Intel X25-M G2 80GB + Hitachi 7K1000.C 1TB
- PSU: PC Power & Cooling Silencer 910W 80+ Silver
- Case: Thermaltake Level 10 GT

All Catalyst driver is coupled with their corresponding, then-newest Catalyst Applications Profile (CAP) - that is, for example, CAP5 for 11.3, CAP2 for 11.4.

1. Aliens vs. Predator 3 - Benchmark tool
(DX11 / Texture Quality: Very High / Shadow Quality: High / Anisotropic Filter: 16x / SSAO On)

▲ Result summary for this game is as follow:
- Catalyst 11.2 does not support 3-way CFX profile.
- Catalyst 11.4 & 11.5 decline MLAA performance / decrease CFX efficiency when MLAA is applied.
- (Qualitiative) Rapidly-moving scene is over-blurred when MLAA is applied, for Catalyst 11.4 & 11.5

CFX scale in this game is pretty good:
- 2-way 197~200% in average (regardless of Catalyst version)
- 3-way 291~300% in average (except 11.2 which doesn't support 3-way profile for this game)
But CFX scale is decreased slightly when MLAA is applied under Catalyst 11.4 & 11.5:
- 2-way 180~190%
- 3-way around 260%

Following is single VGA performance comparison.


▲ No significant change is observed except MLAA failure under 11.4 & 11.5.
I guess morphological filtering algorithm is changed since Catalyst 11.4, but not sure.


2. Battlefield: Bad Company 2 - In-game test (Operation Aurora)
(DX11 / High preset + HBAO On)


▲ Result summary for this game is as follow:
- Catalyst 11.4 & 11.5 decline MLAA performance / decrease CFX efficiency when MLAA is applied.

The graph seems like framerate is bottlenecked around 110~120 by some other than VGAs - CPU perhaps.
CFX scale goes up when anti-aliasing is applied due to CPU bottleneck, and the scale under AA looks fine.
- No AA: 2-way 176~183% / 3-way 185~195%
- 4x MSAA: 2-way 192~200% / 3-way 226~235%
By the way, CFX scale under MLAA (before vs. after 11.4) seems interesting:
- ~11.3: 2-way 198~199% / 3-way 278~280%
- 11.4~: 2-way 186~187% / 3-way 261~262%
See what I mean?

Here's single VGA performance for this game.


▲ Regardless of problems seen above for multi VGA configuration, single VGA performance goes up as Catalyst is renewed - except for MLAA performance deficit under 11.4 & 11.5.


3. Crysis: Warhead - Benchmark tool (Avalanche)
(DX10 / Enthusiast preset)


▲ Result summary is as follow:
- Catalyst 11.4 & 11.5 decline MLAA performance / decrease CFX efficiency when MLAA is applied.
- (In addition to above) Catalyst 11.5 decreases CFX efficiency 'a bit' in general (regardless of AA settings).

It looks clear that Crysis: Warhead scales nicely with multi-VGA, both for 2-way and 3-way.
CFX scale is relatively 'homogeneous' regardless of Catalyst version or AA settings, but Catalyst 11.5
- ~11.4: 2-way 185~191% / 3-way 281~287%
- 11.5: 2-way 183~188% / 3-way 270~276%

Here's single VGA performance comparison for this game.


▲ Again, whatever you saw just before for multi-VGA graph, single VGA performance is increased as Catalyst is renewed - and poor MLAA performance under 11.4 & 11.5 is also observed.


4. Crysis 2 - In-game test (Second chance)
(Extreme preset + SSAO On by editing cfg file)


▲ Result summary is as follow:
- Catalyst 11.4 & 11.5 decrease CFX efficiency drastically
- (In addition to above) Catalyst 11.4 & 11.5 decrease CFX efficiency when MLAA is applied (= CFX scale for MLAA is poor)
- Catalyst 11.2 & 11.3 decrease CFX efficiency when MSAA is applied.
Note that Crysis 2 is currently unable to set MSAA or SSAA in in-game control panel, so I needed to edit cfg file again for MSAA test.

It's weird that CFX scale under 11.4 & 11.5 falls drastically - one thing awe me the most is, nevertheless CFX scale against single VGA falls, 2-way : 3-way performance ratio is still araound 1 : 1.5. It means some 'static' GPU usage loss is occured once multi-VGA is detected. Is it for cutting Watts? I'm not sure.
For detail, GPU usage under 11.4 & 11.5 at multi-VGA configuration seems topped around 75~80% so that conceptually CFX scale is made as follow:
- single: 100%
- 2-way: 80% + 80% = 160% (assume that, w/o concerning GPU usage, CFX scale itself is 200%)
- 3-way: 80% + 80% + 80% = 240% (assume that, w/o concerning GPU usage, CFX scale itself is 300%)
Now, I think it would explain a lot for the result showed above.

And there's another strange result in this game: 11.2 & 11.3' poor CFX scale for MSAA - I've tried, but I cannot find any possible hypothesis for this phenomenon. One possible (trivial) assumption is that, up to Catalyst 11.3, Crysis 2 is not fully supported.

Here's single VGA performance comparison for this game.


▲ Again, single VGA performance is 'slightly' increased as Catalyst is renewed. It's also the first (and last, in advance) result without MLAA performance deficit under 11.4 & 11.5.


5. Metro 2033 - Benchmark tool (Frontline)
(DX11 or DX10 / Quality: Very High / Anisotropic Filter: 16x / PhysX Off)


▲ Only for this game, I tested also two inferior settings together with the highest setting (DX11 / 2560 x 1600):
- DX11 / 1920 x 1200
- DX10 / 2560 x 1600
The reason is simple: since the game is highly hardware demanding, framerate under the highest setting is terribly low. I need some 'meaningful' framerate not only for 3-way CFX but also for single VGA. Too 'small quantity' of framerate in compare with the minimum decimal significant increases the width of distribution chart - means bigger standard deviation.

In addition to above, I also disabled Advanced PhysX because it is CPU demanding instead of VGA, at least under AMD's GPU.

Result summary for this game is as follow:
- Catalyst 11.4 & 11.5 decline MLAA performance.
- Catalyst 11.5 decreases CFX efficiency a bit.

Under DX11. CFX scales among AA settings are indifferent (since, even without AA, the framerate is very low - which means it's highly VGA demanded) and the scale seems not bad.
- (DX11 / 2560) 2-way around 190% / 3-way 250~260%
- (DX11 / 1920) 2-way around 180% / 3-way 220~230%
But for DX10, CFX scales start to differ by AA setting.
- (DX10 / 2560 / No AA) 2-way 178~180% / 3-way 206~208%
- (DX10 / 2560 / with AA) 2-way 185~190% / 3-way 230~240%
Note that for 4x MSAA and MLAA CFX scales are indifferent (it doesn't mean their performance is necessarily the same).

Here's single VGA performance for this game.


▲ Single VGA performance goes up as Catalyst is renewed, and MLAA performance falls by 20% after Catalyst 11.4 - same patterns as before.


6. S.T.A.L.K.E.R: Call of Pripyat - Benchmark tool (Sun shafts)
(DX11 / Ultra preset + HDAO Ultra as SSAO option)


▲ Result summary for this game is as follow:
- CFX scale goes worse as Catalyst is renewed.
- (In addition to above) Catalyst 11.4 & 11.5 decline MLAA performance.

CFX scales among different AA settings are indifferent. CFX scales are as follow:
- 2-way: 186~195% / 3-way: 250~274% - It seems appropriate.

Here's single VGA performance for this game.


▲ Finally we found a counterexample for previous results: the performance goes worse as Catalyst is update! :-(


Up to here, the whole results can be summarized as below:

- Single VGA performance is increased as Catalyst is update, in general (except for SCoP).
- MLAA performance & CFX scale is worsen after Catalyst 11.4 WHQL (guess some algorithm change is inserted)
- CFX scale for Crysis 2 is drastically drop after Catalyst 11.4 WHQL


My suggestion is:

- For multi-VGA users, up to now Catalyst 11.3 shows the best performance in average.
- For single VGA users, Catalyst 11.5 (and perhaps then-newest drivers ever after) is the best.


So that's the end of this posting. Thanks for reading my article :-)