Pixel and Vertex Shading Performance

Pixel and Vertex Shading Performance

How did Pixel and Vertex Shading performance change during the last four months? We looked for an answer to this question by using previous GPU tests.

Here are Pixel Shading results obtained with our test using pixel shader 2.0 extracts from standard applications and processed with another of our own applications (results are in Mpixels/s):


ATI’s Radeon X800 XT PE is still the fastest. Results are more complicated for the X800 Pro as the 6800GT is faster for two of the three shaders. The 6800GT processes the pixels in FP32 instead of FP54 for ATI. But the most interesting is the performance improvement brought by NVIDIA to their drivers; the 3DMark03 shader is 20% faster and 8% faster for Tomb Raider: Angel of Darkness.

As these shaders are processed with our application, there are no specific optimizations. It’s possible to imagine that optimizations are generic and are located on the NVIDIA’s shader compiler better at using the GeForce 6 architecture. This performance improvement isn’t verified with the GeForce FX, and the Radeon’s performance is unchanged.

The next test results involve the Vertex Shader. For this test we used a RightMark 3D fixed function geometrical test (T&L), in VS .1, VS2.x and in VS3.0. The same operation is calculated each time.


NVIDIA keeps its advantage for T&L operations. NVIDIA’s vertex units seem strongly optimised for this mode. With vertex shaders ATI had the best results. However, the performance gap is small with Vertex Shader 1.1 and 2.0. NVIDIA reduced this gap with the new drivers. The performance gain is close to 7% since our last tests.

As with the previous test NVIDIA’s GPU performance drops with static branching, but less than before. Performance was improved by 28%! Dynamic branching is still carried out slightly faster. It’s not completly logical, but by looking at the performance gain it’s more than encouraging.

Filtering optimizations

Filtering optimizations

NVIDIA and ATI started improving their filtering process a few months ago. Filtering requires a lot of resources. Trilinear filtering requires twice as much as bilinear and 8X anisotropic filtering requires 8 times more resources than trilinear filtering! In such situations it’s interesting to find a good compromise between picture quality and reduced load.


From the best to the worse... what is the right compromise.
The problem with compromises is that they are more or less hidden by the manufacturer. At the time of the GeForce FX, we saw that NVIDIA used more “aggressive” compromises to improve performance when detecting games like Unreal Tournament. Recently, ATI tried to be smarter by integrating a compromise between bilinear and trilinear to the X800 like NVIDIA did. Because this algorithm was not visible in verifying quality tests for colored mipmaps, they denied doing it.

Now after being discredited, NVIDIA allows activation or deactivation of these optimizations via the drivers, except angle anisotropic optimization on the GeForce6. ATI also announced that they would do the same thing four months ago after being caught. We are still waiting…


NVIDIA´s drivers permit the optimizations deactivations
ATI deserves criticism for this problem and NVIDIA should be recognized for their initiative, even if there is still a lack of documentation on optimizations to make everything perfect. At any rate optimizations are not a bad thing for the user. NVIDIA’s performance with Unreal Tournament 2004 in 1600*1200, using significant texturing and filtering, is strongly improved:

Optimisations deactivated (High quality)
GeForce FX 5950 Ultra : 37.6 fps
GeForce 6800 GT : 74.7 fps

Optimisations deactivated (Quality, Trilinear optimization On, Anisotropic optimisation On)
GeForce FX 5950 Ultra : 52.1 fps (+38.6%)
GeForce 6800 GT : 91.9 fps (+23.0%)

Most users will be pleased with this performance gain. If you think that quality is reduced, it’s possible to deactivate all or a part of these optimizations. We are sorry it’s not possible to do this with ATI without getting lost in the register base, but we hope this situation is only temporary.

Anti aliasing and Anisotropic filtering

Anti aliasing and Anisotropic filtering

We already talked about graphic quality in past articles, mainly anti aliasing and anisotropic filtering. Just to review, anti aliasing provided by ATI’s chips is superior to the GeForce FX’s, but the GeForce 6800’s is relatively close. In terms of anisotropic and bilinear filtering without optimizations, NVIDIA has better results in synthetic tests designed to put this capability in value. However in practice differences are small.

For a better idea, here are a couple of pictures extracted from UT 2004, Colin McRae 04 and Far Cry. Picture quality is 1600*1200 with 4X anti aliasing and 8X anisotropic filtering forced by the driver (the pictures’ original version non compressed- except for Far Cry’s- are downloadable here). Of course these are fixed images and it’s only possible to evaluate what the overall quality of a picture in motion would be. They are in the following order: GeForce FX, Radeon 9800, GeForce 6800, Radeon X800 :




Here is the impact on performance of anti aliasing 4X and anisotropic filtering for two of the three games (Far Cry has performance restricted by the CPU, so results are not admissible). Anisotropic filtering is forced via the drivers and optimisations are activated:


With anti aliasing 4X, the GeForce FX clearly has the best results. In 1600 x 1200 4X, it’s possible to reach on average 80% of performance with the two games. The next graphic cards are the GeForce 6800 GT and X800 Pro at 73%, the 9800XT at 71% and the X800 XT PE is last with 65%.


The GeForce FX 5950 doesn’t provide equivalent results with anisotropic filtering. Despite the driver’s optimisations this graphic card is last with only 68% of performance in 1600*1200 Anisotropic 8x. The Radeon 9800 has a comparable score, which is less than the X800, because of the lack of trilinear optimisation of the previous ATI chips. Ahead of the 6800, the X800 provides the best performance with an honourable 80%.


Finally, when the 4X anti aliasing and 8X anisotropic filtering are activated, results are equivalent. And by looking at the previous results this is quite logical. The X800 suffers the least followed by the 5950U, the 6800GT, the X800XT PE and finally, the 9800XT.

MSI NX800 & RX800, test protocol

Test configuration

For this test we used the following configuration:

- ASUSTeK P4C800-E Deluxe
- Intel Pentium 4 3.4E GHz
- 2x512 MB DDR PC3200 in 2-3-3-8
- Windows XP SP1 French
- DirectX 9.0c

The latest updated drivers officially available were used, the ForceWare 61.77 for NVIDIA and the Catalyst 4.8 and 4.9b (for Doom III) for ATI. We tested the following graphic cards; Radeon 9800 XT, Radeon X800 Pro, Radeon X800 XT PE, GeForce FX 5950 Ultra, GeForce 6800 GT and GeForce 6800 Ultra. In order to compare we also integrated the GeForce 6800 for its standard performance.

If the GeForce 6800 Ultra and ATI Radeon X800 XT PE are two cards of reference for NVIDIA and ATI, the GeForce 6800 GT and ATI Radeon X800 Pro aren’t. Both are manufactured by MSI.


These two graphic cards come with three games (Prince of Persia the Sands of Time, Uru Ages Beyond Myst, XIII), InterVideo software (WinDVD et WinDVD Creator Plus) for playing and burning DVDs, and a few utilities (Photoshop Album SE, 3D Album, RestoreiIT and VirtualDrive). There is also an S-Video cable and a DVI to VGA adaptor in the package.

MSI chooses to follow the almost perfect ATI design for the RX800Pro, but changed NVIDIA’s 6800 GT cooling system by replacing it with its own solution. The copper radiator’s fan speed can be adjusted from 2800 RPM to 4000 RPM, which is useful for overclocking. In practice, the NX6800GT is the noisiest in stand-by or in normal use while the RX800 Pro is the quietest. However, under intensive use the RX800 becomes the noisiest. Fan speed increases according to ATI’s high standard design. But for both graphic cards the noise remains acceptable.

Test protocol

Anistropic filtering was activated almost every time via the driver. All optimisations were activated for ATI (including bilinear filtering for level 1 texture) but also for NVIDIA to get comparable results. Only Doom III was the exception. Doom III’s internal parameters are very precise regarding anisotropic filtering and are only activated for the textures that require it. However, these parameters are deactivated when anisotropic filtering is forced via the drivers and strongly reduces performance for a small improvement in graphic quality.

Anti-aliasing was also activated via the drivers, but activation via the driver or game doesn’t change performance or quality. The exception was Far Cry. NVIDIA’s anti-aliasing doesn’t change anything for this game. It has to be deactivated because of performance issues.

Anti-aliasing has to be activated via the game or else the comparison is made between ATI with anti aliasing and NVIDIA without. This situation would clearly be an advantage for NVIDIA, and unfortunately, some journalists have already made this mistake.

We changed a couple of games from our test protocol. Unreal Tournament 2003 was replaced by Unreal Tournament 2004. We used our own CTF demonstration with one of the new maps of the game to measure graphic card performance.

Far Cry is still in the test but we now use the 1.2 patch. Some may think this sacrilegious, because it was discarded by the Crytek team. However, none of our tests showed any noticeable problems. Graphically, it’s even better. Contrary to the 1.1 patch, which shows some quality difference regarding the different shaders, the 1.2 patch GeForce path is less aggressive on the shader level in partial precision.

The 1.2 patch provides the possibility to activate shaders 2.0B or 3.0 for the X800 and 6800 and use longer shaders able to process 3 to 4 dynamic lights in one path (3 for the 2.0B, 4 for the 3.0) instead of 1 for the standard 2.0 shader. As this type of light is more present indoors, we modified our test scene. This scene is in a new map. 2/3 of the time spent in this scene is outdoors and 1/3 indoors, a ratio representative for us of the time spent inside/outside in Far Cry.

Tomb Raider : Angel Of Darkness was replaced by the latest FPS release, Doom III. We didn’t use the demonstration scene included in the game for our tests in order to avoid specific driver optimizations. We used our own saved scene.

IL-2 Sturmovik Forgotten Battles is still in our test but we moved into second gear with the Ace Expansion Pack extension. Shader 2.0 is now available for the water if the Perfect / Maximal graphic resolution is selected. This option is a heavy load, so we made the tests with and without this option. The saved scene was provided by Michel Vibert.

We still used Splinter Cell : Pandora Tomorrow, Warcraft III, Colin Mc Rae Rally 04 and Fifa 2004 with a few modifications to facilitate our benchmarks. Half Life 2 will be included as soon as it’s released as we didn’t think that including Counter Strike : Source Beta in benchmark mode for this test was appropriate. When available a scene of our choosing from Half Life 2 will be incorporated into future tests.

UT2004, Far Cry

Unreal Tournament 2004

We start with Unreal Tournament 2004. With the top-of-the-line graphic cards performance was restricted by the internal processor in 1600*1200 or in 1024*768 with AA 4x / Aniso 8x. This explains the small gap between the X800 Pro / X800 XT and 6800 GT / 6800 Ultra with these resolutions.

With any parameters, the 6800GT is slightly faster than the X800 Pro, and the 6800 Ultra is also a little more efficient than the X800 XT. With the higher end graphic cards, the fps reached in 1600*1200 AA 4x / Aniso 8x is equivalent to the previous higher end graphic cards fps in 1024*768 AA 4x / Aniso 8x. The card of reference, the GeForce 6800, has better results than the FX 5950U / 9800XT, except in 1600*1200 4x/8x because of a 128 MB video memory instead of a 256 Mb for other graphic cards.

Far Cry

For Far Cry we used the SM2.0B or SM3.0 path when possible. This path makes possible the use of longer shaders to accelerate the management of dynamic lightning, and also Geometry Instancing for the management of grass in an outdoor setting.

In practice, performance gains reached 20% in specific situations. However, in our demonstration scene, meant to be the most representative of the game (2/3 outside, 1/3 inside), performance gains were more limited. 4.1% with a Radeon X800 XT PE in 1600*1200 AA 4x / Aniso 8x and 8.2% with a GeForce 6800 Ultra was the improvement measured after activating the SM2.0B path.


For Far Cry the Radeon X800 XT PE is clearly the fastest, followed by the GeForce 6800 Ultra, and the Radeon X800 Pro and GeForce 6800 GT are very close, one better than the other at times depending on the resolution. The GeForce FX 5950U didn’t provide great results and was clearly slower than the GeForce 6800. The 6800’s performance, however, wasn’t excellent and was just a little better than the Radeon 9800XT.

Doom 3, Splinter Cell : PT

Doom 3

Doom 3 uses the latest graphic engine designed by ID Software and John Carmack. Using OpenGL, it overwhelmingly uses volumetric shadows. These types of operations are easier for NVIDIA’s GPU, which is able to process twice as much Z data per cycle than ATI’s VPU without FSAA. Doom III was tested in “Very High” with the 8X mode anisotropic filtering activated.

ATI, aware of the performances issues with Doom 3, has released a Catalyst 4.9 pre-version supposed to improve game performance. We used this version because it was available on ATI’s website. Here are performance gains obtained compared to Catalyst 4.8:


The figures in the table show that improvements with the Radeon 9800XT are inexistent, small with the X800 Pro and significant with the X800 XT PE. We hope that the final Catalyst 4.9 version will improve the X800 Pro’s performance like the pre-version did with the X800 XT.


Basic Radeon X800 performances are far behind. If the GeForce FX 5950U is only slightly ahead of the Radeon 9800 XT, the Radeon X800 XT PE is between the 6800 and the 6800GT. The Radeon X800 Pro doesn’t really benefit from the 4.9 pre-version and is clearly behind the X800 XT PE. Activating the anti-aliasing helps the X800 XT PE get closer to the 6800GT, even if it remains clearly behind. We noticed the 2.2 fps provided by the GeForce 6800 with the highest settings: 1600*1200 AA 4x / Aniso 8x. A 128 MB graphic card isn’t powerful enough for Doom III in Very High resolution. In lesser settings, the performances were excellent, superior to the X800 Pro and clearly ahead of the 5950U or the 9800XT.

Splinter Cell : Pandora Tomorrow

After having a lot of trouble with Doom III, the Radeon cards are back on track thanks to Splinter Cell. The X800 XT PE provides the best results and has a clear advantage on the 6800 Ultra. The X800 Pro is between the 6800GT and 6800Ultra.

As for the 6800, and previous top-of-the-line, it has better performances. The 5950U and 9800XT provide almost equal results.

IL-2 FB AEP, Warcraft III

IL-2 Sturmovik Forgotten Battles Aces Expansion Pack

Here as promised, two benchmark series with a new add on, the new map and a new demonstration scene provided by Michel Vibert, whom we thank once more.


With the landscape in “excellent” mode, the IL2 standard setting for simplified highest setting, the results are in the usual order: the Radeon X800 XT PE, the GeForce 6800 Ultra, the GeForce 6800 GT, and finally, the Radeon X800 Pro.

Between the 6800, 9800XT and 5950U, the 5950U provides the worst results. The 6800 takes the lead, ahead of the 9800XT, from 1/10th to 12.8 fps depending on the settings.


When advanced settings are pushed to the limit with the landscape details in “perfect”, IL-2 activates –amongst other parameters – the Shader 2 for the water. The results are bewildering but costly in terms of performance:


In 1600 x 1200 pixels, the X800 XT PE is still in the lead followed by the 6800 Ultra. The X800 Pro is a little late with a performance between the 6800 and 6800 GT. The GeForce FX 5950U isn’t comfortable at all in this mode, and performance drops 69% compared to the previous figures!

Activating the AA 4x and the Aniso 8x changed the results. NVIDIA’s 6800GT take the lead ahead of the X800 XT PE. The X800 Pro didn’t provide great results either. The card of reference, the GeForce 6800, is faster in 1024. ATI’s chips hold the lead in 1600 but the framerate is too low to have nice gameplay with this resolution. The 5950U has a performance close to the 9800 XT.

Warcraft III

Warcarft III is a typical example of a performance restricted by the internal processor. Even if any chosen moment in the game replay is graphically heavy, you will notice that with the X800 XT or the 6800U changing from 1600 x 1200 to 1600 x 1200 AA 4x / Aniso 8X doesn’t significantly change the level of performance. The system can’t provide more than 54 fps to the graphic card. The GeForce FX 5950U is definitively last in this test, and the standard 6800 is slightly in front of the 9800 XT.

Colin Mc Rae Rally 04, Fifa 2004

Colin Mc Rae Rally 04

We now come to testing car simulations with the Colin Mc Rae Rally 2004 version. The Radeon X800 XT PE is the most efficient with this type of game in front of the 6800 Ultra. The X800 Pro is slightly behind the 6800 GT.

The standard 6800 is far more efficient than the GeForce FX 5950U. The GeForce FX 5950U is like all GeForce FXs not comfortable with car simulations.

Fifa 2004

The last test is Fifa 2004. Here ATI’s VPU clearly gives the best results even if the framerate provided by most of the graphic cards is enough to play in very good conditions. We notice that in 1600*1200 AA 4x / Aniso 8x, the le X800 Pro provides better results than the 6800 Ultra. The Radeon 9800 is between the 5950U and the 6800.

Overall performances, Viewperf 8.0

Overall performances

In order to have a general view of performance, we normalized the results for each test giving the best result 100%. We then accumulated the scores to compare the cards based on the Radeon 9800 XT’s 100% performance.


For the previous top-of-the-line products, the Radeon 9800XT confirms its supremacy once more over the GeForce FX 5950 Ultra. Regarding the current top-of-the-line products, results are simple to analyze. The Radeon X800 XT PE is first just in front of the GeForce 6800 Ultra in 1600 x 1200 4x/8x. A little behind is the 6800 GT followed by the X800 Pro.

In 1600 x 1200 32 bits 4X/8X, without Doom 3, the figures obtained are: X800 XT 202%, 6800 Ultra 188%, 6800 GT 170%, X800 Pro 164%, 6800 110%, 9800 XT 100% and 5950U 86%.

The superior evolution of NVIDIA’s drivers compared to ATI’s and the changes made to our test protocol changed the GeForce 6800’s results. These graphic cards are much more comfortable now in tests then 4 months ago. In the first tests, the X800 XT PE was clearly in the lead and the X800 Pro’s performance was comparable to the GeForce 6800 Ultra! The situation is now different and NVIDIA has benefited. This may change, however, with the release of Half life 2 or with another ATI driver version for Doom 3.

Viewperf 8

Here are the results of the tests made with Viewperf for every graphic card. Viewperf is used to measure the OpenGL « Pro » performance by simulating typical resources required by professional software.

- 3ds max
- CATIA
- EnSight
- Lightscape
- Maya
- Pro/ENGINEER
- SolidWorks
- Unigraphics


Of course, unlike the FireGl or Quadro, these graphic cards are not meant to be used with this type of software. The FireGl and Quadro are based on the same hardware but have specific drivers. Buying this type of graphic card isn’t possible for everyone, but this test remains interesting:


The 6800 is ahead of the X800 in 3ds max, EnSight, Maya, ProENGINEER and SolidWorks. With CATIA, Lightscape and Unigraphics results change and the X800 is the fastest.

Conclusion

Conclusion

The graphic card market is constantly changing. NVIDIA’s GeForce 6800 shows more improvement than the X800. This result isn’t surprising as mentioned in the conclusion of the X800 article.

ATI keeps the performance leadership thanks to the Radeon X800 XT PE. The Radeon X800 XT PE is more efficient than the 6800 in processing shaders and overall is slightly faster in high resolution with anti aliasing and anisotropic filtering in games (with exception, of course, to Doom 3).

As for the GeForce 6800 GT and X800 Pro the best results come from NVIDIA’s solution.

It seems however that NVIDIA and ATI have trouble in producing a large number of their top-of-the-line chips. This situation will change in the upcoming weeks, but there is another question to be asked. Are these supply issues related to the difficulty in producing high-end chips? And will graphic card evolution slow down because of this problem?