Connect3D Radeon 9700 Pro in the test: performance leap at high resolutions

Connect3D Radeon 9700 Pro in the test: performance leap at high resolutions

Table of contents
  1. 1 Introduction
  2. 2 The card
  3. Scope of delivery
  4. 3 Drivers
  5. 4 Specs and technology
    1. 5 FSAA and AF
  6. 6 Benchmarks
  7. Test system
  8. 7 Synthetic tests
    1. 3DMark2001 SE
    2. 8 3DMark Detail
    3. 9 Villagemark D3D
    4. 10 Villagemark Detail
    5. 11 TempleMark D3D
    6. TempleMark D3D detail
    7. 12 Codecreatures Bench
    8. 13 Codecreatures Detail
    9. 14 Aquamark D3D
    10. 15 Aquamark D3D detail
  9. 16 Game benchmarks
    1. Ultima IX D3D
    2. 17 Ultim a IX D3D Detail
    3. 18 Comanche4 D3D
    4. 19 Comanche4 D3D Detail
    5. 20 Dungeon Siege
    6. 21 Dungeon Siege Detail
    7. 22 Aquanox
    8. 23 Aquanox in detail
    9. 24 Jedi Knight II
    10. 25 Jedi Knight II Detail
    11. 26 Max Payne
    12. 27 Max Payne detail
    13. 28 Serious Sam SE D3D
    14. 29 SeSam SE D3D Detail
    15. 30 Serious Sam SE OGL
    16. 31 SeSam SE OGL Detail
    17. 32 UT2003 Demo FlyBy
    18. UT2003 Demo Botmatch
    19. 33 UT2003 FlyBy Detail
    20. UT2003 Botmatch Detail
    21. 34 Alice
    22. 35 Alice Detail
  10. 36 Conclusion

TempleMark D3D

The Templemark is just like that previous Villagemark, originally a demonstration program from PowerVR. However, since it supports a lot of current features such as hardware TnL, bump mapping and up to 6 texture layers in one pass, it is also well suited as an independent benchmark that is guaranteed not to be optimized for nVidia chips.

TempleMark D3D v1.06 -default
  • 1024x768x32 :
    • R9700pro (default)
    • GF4 Ti4600 (default)

It's relatively tight, only good 11fps, or just under 8 The Radeon9700pro can distance itself by 5% from the Ti4600, which corresponds almost exactly to the 8.3% clock difference in favor of the 9700pro. The slightly lower overdraw than in the village mark, the up to six texture layers and the type of texture operations (bump mapping, environment mapping) do not suit the Radeon9700pro so well. In spite of all this, it can assert itself well and clearly wins, but the multitexturing ability of the chip is more in demand than the bandwidth. With its 6x single pass texturing, the Radeon8500 was able to present itself very well here, even if it was just beaten by the GeForce4. In situations like this, the design decision to continue to trust only 8 texture units and 'only' split them up differently (and more sensibly).

Let's take a look at the quality settings and let us be surprised whether the Radeon9700pro can pull away miles here ...

TempleMark D3D v1.06 -Quality
  • 1024x768x32:
    • R9700pro (2xAA/4xAF)
    • R9700pro (4xAA/8xAF)
    • R9700pro (6xAA/16xAF)
    • GF4 Ti4600 (2xAA/4xAF)
    • GF4 Ti4600 (4xAA/8xAF)
    • GF4 Ti4600 (4xS-AA/8xAF)

... Yes, it can.

Despite similar performance in the raw output, i.e. without image quality-improving features, the Radeon is much more careful with this raw output and is able to maintain a large part of it with increased image quality. The victory is clear once again, even if the gaps are not quite as gigantic as in the Villagemark. Despite the optically quite linear diagrams, the Radeon demonstrates the GeForce4 Ti4600 with the same settings with double to triple performance.

TempleMark D3D Detail

How much of the performance drop with the quality settings is the lack of memory bandwidth and/or not due to inadequately designed hardware at the FSAA?

TempleMark D3D v1.06 -FSAA
  • 1024x768x32:
    • R9700pro (default)
    • GF4 Ti4600 (default)
    • R9700pro (2xAA)
    • R9700pro (4xAA)
    • GF4 Ti4600 (2xAA)
    • R9700pro (6xAA)
    • GF4 Ti4600 (4xAA)
    • GF4 Ti4600 (4xS-AA)

Here you can clearly see that nVidia has limited itself to fast 2xFSAA when designing the GeForce4 and that the 4x mode was optimized on the side. The Radeon, on the other hand, is obviously a pure-bred FSAA creature, because the highly optimized FSAA of the card can be displayed within the framework of the almost doubled memory bandwidth compared to a GeForce4 Ti4600 without any further loss of performance. Even with 6xFSAA, the card hardly falls off, while the GeForce buckles significantly with more than twice the FSAA.

TempleMark D3D v1.06 -AF
  • 1024x768x32:
    • R9700pro (default)
    • R9700pro (2xAF trilinear)
    • R9700pro (16xAF bilinear)
    • GF4 Ti4600 (default)
    • R9700pro (4xAF trilinear)
    • R9700pro (8xAF trilinear)
    • R9700pro (16xAF trilinear)
    • GF4 Ti4600 (2xAF trilinear)
    • GF4 Ti4600 (4xAF trilinear)
    • GF4 Ti4600 (8xAF trilinear)

With the anisotropic filter, there is the already familiar picture: The qualitatively in principle higher quality method nVidias, almost all textures with the highest possible filter level, regardless of whether it is the really decisive oneBase texture or a lightmap that does not make sense to filter anisotropically. leads to a drop in performance which, without additional optimization by a tweaker like the RivaTuner, makes hardly more than 2xAF appear useful.

All AF modes of the Radeon remain within striking distance of the raw performance of the GeForce4.

On the next page: Codecreatures Bench