VisionTek Xtasy 9800 PRO Page 2

Back to Page 1

Test Setup and Benchmarking Software

The system used for testing the Xtasy 9800 Pro was configured as shown below:

1. Intel Pentium 4 2.4B GHz CPU
2. ASUS P4T533-C motherboard (533 MHz FSB/AGP 4X)
3. 512MB PC1066 RDRAM
4. Xtasy Radeon 9800 Pro using Catalyst 3.4 drivers
5. Windows XP Professional, SP1

The following software was used for benchmarking:

1. Commanche 4 benchmarking demo
2. Splinter Cell (v1.2)
3. Ghost Recon (v1.1.5)
4. F1 2002
5. IL-2: Forgotten Battles
6. Falcon 4 (SP3)
7. 3DMark2003 (330 patch)
8. Codecreatures Benchmark Pro

Unless otherwise noted, all games were tested at their highest settings, with 32-bit color and textures enabled, and V-Sync disabled. Trilinear texture filtering was also enabled by default and all anisotropic texture filtering testing was conducted at Quality (trilinear) settings.

Benchmark Scores and Analysis

Commanche 4

From these scores, it is immediately clear that the game is extremely CPU-limited. Only at the highest setting does the game exhibit any substantial performance loss, and that is most likely due to the frame buffer exceeding the amount of available RAM needed for textures, thereby forcing texture data to be moved across the AGP bus.

Improved texture filtering pretty much exhibits the same pattern as enabling anti-aliasing, with the system processor again acting as the bottleneck.

The most interesting score for combined AA/AF testing was the last one ran at 1600×1200 with 6xAA/16AF enabled; in fact, this is the only setting that produced an unplayable frame rate and I’m curious to see whether or not a 256MB board would make a real difference here. Regardless the game is so CPU-limited that its value as a graphics benchmark is questionable.

Splinter Cell

The 1.2 patch for this game made available a benchmarking tool for testing the title’s performance. And though the patch also included three demos, they appear to be too dependent on CPU performance to be good graphics tests so I’ve opted to instead use the Beyond3D Caspian Oil Refinery demo (which can be found here).

Higher resolutions certainly impact performance, with 1600×1200 resulting in a 25% loss of frame rate, though improved texture filtering had little affect. This is most likely due to the texture-sparse nature of the demo itself, with its emphasis on shader performance due to the shaded water and brief use of goggles. Also noteworthy are the identical scores at 1280×1024, which nothing I did would change; I definitely plan on keeping an eye on this peculiarity in future reviews. Last, due to Splinter Cell’s lighting system causing graphical anomalies in the game with anti-aliasing enabled, the game obviously could not be tested with this feature.

Ghost Recon

I tested Ghost Recon using the SimHQ benchmark mission (found here).

Ghost Recon is a game that continues to scale quite well as graphics boards become more powerful. Even with 6x anti-aliasing enabled, the frame rates stay extremely high. Curiously enough, Ghost Recon is the one title tested for this review that would not run at 1600×1200 with 6x AA, though according to ATI this setting is supported by the board; the mission would run just fine, but anti-aliasing had been disabled.

We can see that improved texture filtering does not have nearly the performance impact that anti-aliasing did. Even at the highest settings, the game lost a mere 20% in frame rate.

Even with the settings maxed out, Ghost Recon still remains extremely playable, though 1600×1200 again refused to allow 6x AA to be enabled.

F1 2002

This racing sim was tested using the built-in ISI playback.

Strangely enough, 2x AA seems to incur quite a performance hit, particularly at higher resolutions. The next setting, 4x AA, also greatly reduces the frame rate but the hit is not linear compared to that of 2x, and 6x AA continues this trend of diminishing performance degradation. The game, however, does remain playable at 1600×1200 with 4x AA and looks extremely good.

Anisotropic texture filtering takes very little from this title, but this is somewhat understandable with the game’s repeated track textures sparse roadside details.

2x AA again takes its toll on the frame rate here, with 4x AF incurring roughly another 5% performance hit across the tested resolutions, though all but the highest setting remains fairly playable.

IL-2: Forgotten Battles

This WW2 flight simulation was tested using the SimHQ benchmark mission located here.

Much like Commanche 4, Forgotten Battles would appear to be fairly CPU-limited if you consider that 1280×960 with 6x AA is only 10% slower than 1024×768 with no anti-aliasing at all. The worse frame rate again comes at 1600×1200, with the performance drop indicating that 6x AA isn’t leaving enough storage in the onboard RAM for the required textures.

Higher texture filtering is for all intents and purposes essentially free at the lower two resolutions, and only suffers a 10% frame rate loss at 1600×1200 and 16x AF.

Combined with anisotropic filtering, the anti-aliasing performance is hardly affected at the lower resolutions. Even 1600×1200 maintains playable frame rates until 6x AA is enabled.

Since Forgotten Battles supports both D3D and OpenGL, I decided to test the latter API to see if there are any significant performance differences between the two.

Here we can see that running under OpenGL does offer roughly a 20% performance increase over D3D, though this delta is somewhat eroded to 10% once AA and AF are enabled. OpenGL also allows for the in-game ‘Perfect’ landscape detail setting to be selected, an option that brings the game’s frame rate below its equivalent D3D score by approximately 10%. Future reviews will most likely use this API as the primary test setting for the game.

Falcon 4 (SP3)

This aging, yet still popular, sim was tested using the Papadoc timedemo (found here).

Perhaps due to its age, anti-aliasing takes more of a performance hit in Falcon 4 than with more recent simulations. 2x AA takes roughly a 20% performance hit, save at 1600×1200 where the frames lost are closer to 33%, and 4x and 6x AA are very close in performance across all resolutions. The lower resolutions tested are certainly playable, even with 6x enabled, but 1600×1200 unfortunately dips too far into the low 20s.

While anisotropic filtering makes almost no impact whatsoever on the frame rate, its benefit to the game’s image quality is remarkable. Without AF enabled, ground textures, particularly in the lower resolutions, display an extremely annoying amount of aliasing and moire; once anisotropy is applied the aliasing is almost completely ameliorated.

These tests show frame rates that are quite similar to those with only anti-aliasing enabled. And while 1600×1200 is again probably too slow for most gamers’ preference, Falcon 4 still looks amazing at 1280×960 with 6x AA and 16x AF.

3DMark2003

Though Futuremark’s synthetic benchmark has become a topic of controversy in recent weeks, it still remains a useful tool by which reviewers and gamers may examine DX9 shader performance. While few game titles will have support for these features this year, or perhaps even next, DX9 shader performance remains a key selling point through which IHVs such as ATI hope to promote and sell their latest products. And with game development generally several years behind in its support of the latest hardware features, synthetic benchmarks allow us an early, and hopefully accurate, examination of such cutting-edge features.

This synthetic benchmark was ran with the 330 patch installed. Worth noting is how much the single- and multi-texturing fill rate scores differed, with only the latter even approaching the board’s theoretical peak fill rate. This is most likely due to the board simply lacking the bandwidth, even at 21.8GB, to produce a 32-bit pixel per clock cycle for all eight pipelines.

Last, the Pixel Shader 2.0’s score of 48.2 frames/sec is noteworthy in that no other graphics board currently available comes even close to this score. Whether or not actual game titles using such advanced shaders will also reflect this performance delta remains to be seen.

CodeCreatures

This benchmark is a demo for a game engine that Codecult hopes will attract game developers as licensees. As such, it combines the use of DX 8.1 pixel and vertex shaders with high resolution textures.

Go To Page 3

Related posts:

ATI Radeon HD 4870 Performance Preview Page 3 ATi Radeon X800 XT PCIe Page 5 ATi Radeon X800 XT PCIe Page 6 Gainward GeForce™ 6800 Ultra / 2100 Golden Sample Page 3 Gainward GeForce™ 6800 Ultra / 2100 Golden Sample Page 4 NVIDIA® GeForce™ 6800 GT PCI Express Page 4 VisionTek Xtasy 9800 PRO VisionTek Xtasy 9800 PRO Page 3