3D Image Quality: EMBM, Pixel Shaders, Point Sprites

The correctness and quality of the pixel shaders implementation can be checked with the help of 3D Mark 2001 SE scenes. Remembering about the problems with the texturing quality provided by SiS Xabre, we resorted to SiS.3D.TexTurboMode=3 (it is used by default) and SiS.3D.TexTurboMode=0.

3DMark2001 SE: Nature

SiS.3D.TexTurboMode=0		SiS.3D.TexTurboMode=3
SiS.3D.TexTurboMode=0		SiS.3D.TexTurboMode=3

For a better comparison here is the same scene by NVIDIA GeForce3 Ti200:

As you can see from the screenshots, enabling texture "turbo-mode" works only for textures laid ordinarily: have a look at the bridge stone-work. The quality of the lake surface where the pixel shaders are used doesn't change.

Nevertheless, the level of detail provided by SiS Xabre with Pixel Shaders involved appeared slightly lower here. As you may see, NVIDIA GeForce3 Ti200 creates a "wavier" lake surface:

3DMark2001 SE: Pixel Shader, Advanced Pixel Shader

Pixel Shader by SiS Xabre		Pixel Shader by NVIDIA GeForce3 Ti200
Advanced Pixel Shader by SiS Xabre		Advanced Pixel Shader by NVIDIA GeForce3 Ti200

The only comment here: the surfaces where SiS Xabre uses Pixel Shader look as if their level of detail were slightly lower.

3DMark2001 SE: EMBM, Point Sprites

EMBBM by SiS Xabre		EMBM by NVIDIA GeForce3 Ti200
Point Sprites by SiS Xabre		Point Sprites by NVIDIA GeForce3 Ti200

As for the quality of Point Sprites implementation by SiS Xabre, we have no complaints here. However, EMBM doesn't seem to be its trump at all. If you take a closer look at the screenshot generated by Xabre, you will be able to notice a lot of blocks 2x2 pixels all coloured with one colour. Is it another kind of SiS' own optimization? Well, quite possible…

So, now that we have answered all the question you might have about the image quality, it's high time we passed over to the performance analysis.

Performance

As usual, we would like to start with synthetic benchmarks, as it is much more important than games, you know ;)

Xabre T&L unit demonstrates beautiful performance, which is just a bit lower than that of ATI RADEON 8500.

In case of 8 light sources Xabre takes the lead. SiS' engineers did a great job: since the times of SiS 315 (see our review), when the latter was simply smashed to pieces by the competitor, GeForce2 MX200, the situation has changed drastically :)

Xabre doesn't have hardware support of Vertex Shaders, that is why we were not surprised to see that its result was almost equal to that shown by NVIDIA GeForce4 MX440, as the latter also has no Vertex Shaders unit.

As we have already found out the key responsible for the texturing quality doesn't tell on the pixel shader quality that is why these tests were carried out with the default SiS.3D.TexTurboMode=3:

The hardest tests for SiS Xabre appeared pixel shader performance tests. Of course, it is not shameful to be defeated by a match, but to let your rival turn 2-3 times faster is too much, really…

It is probably the absence of such architectural solutions as HyperZ by ATI and Lightspeed Memory Architecture by NVIDIA, which make the use of the available graphics memory bandwidth more efficient. The second possible reason of this failure might be the imperfect architecture of the Xabre's Pixel Shader unit, because Xabre appeared SiS' first experience in the field.

The fillrate test proves that Xabre doesn't work with the graphics memory so efficiently. This is absolutely evident: working at the same frequencies as ATI RADEON 8500 LE and at much higher frequencies than NVIDIA GeForce3 Ti200, Xabre gets beaten by both.

Only using compressed textures and "turbo-texturing" mode the graphics core managed to show its best and to nearly reach the theoretical maximum:

Now let's pass over to the gaming tests:

Xabre performed excellently in 800x600, however as the resolution grew up it had to step back under the onslaught of more powerful ATI RADEON 8500LE.

The performance drop caused by the disabled "turbo-texturing" appeared not so great as we had expected, and even with the highest quality settings Xabre remained ahead.

Here Xabre performed a bit worse than in car Chase, because there are more textures used and higher Overdraw rates.

In the Quality mode, that is when the image quality is as good as that by the competitors, Xabre falls behind even NVIDIA GeForce4 MX440, while with enabled "turbo-texturing" it managed to reach NVIDIA GeForce3 Ti200.

Here Xabre appeared between NVIDIA GeForce3 Ti200 and GeForce4 MX440 when tested with default settings. However, as soon as we disable "turbo-texturing" it rolls back to the very last position. Luckily, in the intermediate quality mode, Xabre doesn't lose much of the speed.

Nature test uses Pixel Shaders. As we have already seen above, Xabre works with them very slowly. So, no wonder the performance of our hero is not that impressive here.

Despite the beautiful scenes created by the Siege Engine in Dungeon Siege, the textures are not that big and hardly use any multi-texturing. Therefore, Xabre's "turbo-texturing" provides almost no advantage. Here Xabre turned an outsider.

The results of Xabre based cards depend a lot on the texture filtering mode. But in general, Xabre performed very well.

In Quake3 Arena, unlike Dungeon Siege, Xabre had to pay a huge price for the improvement in the texturing quality. Don't be surprised, the results in Quake3 Arena have always been dependent almost only on the texturing speed and graphics memory bandwidth.

Well, I would like to sum up the results of the performance tests and to draw some preliminary conclusions. Having run all the tests and real games we got a very clear idea of "turbo-texturing" and its importance for Xabre family.

Compared with the competitors, Xabre cannot boast very efficient work with the graphics memory. It is most likely to be a not enough "polished" cache structure or their insufficient size. Also please bear in mind that Xabre doesn't have such efficient technologies improving the work with graphics memory as LightSpeed Memory Architecture / LMA II of NVIDIA GeForce3 / GeForce4 MX or HyperZ II of ATI RADEON 8500. Early Z-test presumably implemented in Xabre, which allows the chip to avoid texturing invisible pixels, doesn't seem to be able to save the situation.

In order to make its Xabre baby look nice against the background of more powerful competitors, SiS decided to take another way: they replaced bi-linear texture filtering with approximation, tried to save on selecting additional texture samples, etc. By the way, Xabre can involve its "turbo-mode" not only for speeding up the work with textures, but also for lighting calculations (SiS.3D.LightTurboMode key, disabled by default), for alpha-blending implementation (SiS.3D.AlphaBlendTurboMode key, disabled by default), and for the whole bunch of other functions, I'm sure.

There is nothing bad about these "turbo-modes", actually. If the chip can do something sacrificing image quality for the sake of success, it is excellent, because there will always be users who do not care about clear-cut shapes and correct lighting, but hunt only for performance. On the other hand, I belong to those who prefer beautiful picture to high fps rates.

The conclusion will be as follows: let the users adjust these parameters according to their own needs and everybody will be happy. I hope SiS guys will implement this option in the next driver versions.

Conclusion

Although I am really longing to do it, I will refrain from drawing final conclusions now about the popularity or non-popularity of the new SiS baby.

The chip appeared pretty fast and the cards based on it are most likely to be not very expensive. Pixel Shader support gives Xabre a couple of trumps in the battle with the low-cost competitors from other graphics chip developers. We would also like to stress as one of the indisputable highs SiS' changed attitude towards their products. The drivers have become much more mature and will hopefully grow up to those from NVIDIA and ATI soon. There are almost no artifacts in games that do not fall under any reasonable explanations, and the image quality can be improved within 10 seconds with a couple of simple tricks. The overall picture could be very optimistic for SiS Xabre if…

…if we were sure that the prices of Xabre based graphics cards will be adequate to their performance (with disabled "turbo-modes", of course :) and their functionality. The Value graphics cards market is really packed with the whole lot of products now, and it will be even more packed as soon as ATI RV250 comes out, which will push ATI RADEON 8500/8500 LE into the value sector once and for all. As we have already seen, it is very hard for Xabre to compete with monsters like that, so there will be only one way: to reduce the prices. And how much should a card based on SiS Xabre 400 cost, so that it could attract your attention compared with GeForce3 Ti200 or RADEON 8500/8500LE? I think you can imagine now :)

On the other hand, Xabre competes with the fastest NVIDIA GeForce4 MX, older GeForce2 Pro/Ti and ATI RADEON 7500, which do not support Pixel Shaders. They will also get cheaper and cheaper, that is why Xabre may face some new surprises here as well: its price will have to be attractive against the background of these cards, too.

All in all, what we are looking forward to now, is the finalized drivers (which need a little bit of polishing) and the most important thing: the Xabre pricing info.

Elitegroup, which is well-known as a manufacturer of low-cost solutions of good quality has huge production capacities at its disposal that is why it will do Xabre only good that this company paid special attention to it. ECS AGP400 graphics card based on SiS Xabre 400 made favourable impression having proven reliable and stable throughout the tests. We suppose that the price of ECS AG400 will also be very attractive.

So, let's sum up:

Highs:

• Pretty good performance comparable with that of NVIDIA GeForce4 MX440 and NVIDIA GeForce3 Ti200;
• Pixel Shaders support;
• Very fast T&L unit;
• "Turbo-modes" for some standard functions;
• Multi-monitor configurations, TV-Out and DVI;
• Probably, not very high price.

Lows:

• Low texture filtering quality during "turbo-texturing";
• Sometimes the performance drops significantly when "turbo-texturing" is disabled;
• No multi-sampling support;
• No anisotropic filtering support;
• Dramatic performance drop with FSAA enabled;
• No way to adjust all desired parameters of the chip without addressing the registry keys.