<%BANNER[top_768x90]%>
<%BANNER[banner_468x60_h]%>
<%BANNER[article]%>

Articles: Video

Pages: [ 1 | 2 ]

Performance

3D Performance: Synthetic Benchmarks

Let's start with the most interesting issue: the performance of pixel and vertex shaders by RADEON 8500 and GeForce3 Ti500/200:

In spite of its higher clock frequencies, the pixel shaders unit of RADEON 8500, PIXEL TAPESTRY II, proves to be slower than that of GeForce3 Ti500 in 800x600x16 resolution, though HyperZ II and massive texture caching allow it to break ahead in 1600x1200x32.

The vertex shaders unit, CHARISMA ENGINE II, has demonstrated overwhelming power: its performance was more than twice as high as that of GeForce3 Ti500. As for software acceleration, RADEON 8500 fell a bit behind, because of the absence of FastWrites support as well as of insufficient optimization of geometrical data transfer via AGP.

High Polygon Count benchmark proves once again the complete failure of GeForce3 T&L unit.

It is noteworthy that the twice or even three times as high performance of RADEON 8500 in geometry processing tasks will allow enabling TRUFORM in future games very easily even without introducing any high level of tessellation. In other words, you will have no concerns about the gaming performance and will simply enjoy super high-quality images in the game.

Now come the fillrate tests:

We'd like to remind you that this test is very extraordinarily arranged: the thing we see on the screen is actually 64 semi-transparent textures laid over the polygons as big as the screen. In case of multi-texturing, the polygons are covered with the maximum number of textures the accelerator can handle. Note that in this case the number of polygon "layers" will equal to 64/4=16 in case the accelerator can lay, for example, 4 textures in a single pass. If there's no multi-texturing enabled, 64 polygon layers with a single texture per layer are created.

Returning to our results, we discovered a very interesting phenomenon: although in 16bit mode RADEON 8500 managed to leave GeForce3 Ti500 far behind, it suffered a total defeat in 32bit mode without multi-texturing. Why? We had to run additional tests changing the texture color depth, frame buffer and Z-buffer from 16bit to 32bit. Here is what we've got:

Now the situation has become somewhat clearer: it turns out 32bit frame buffer imposes a great strain over the accelerator thus slowing it down significantly. Probably, RADEON 8500 with its perfect texture-caching system and Z-buffer optimization, draws back only when the amount of information to be read from and written into the frame buffer, which can hardly be cached (though there's an example of such caching: tile architecture of KYRO and KYRO II). The test with disabled multi-texturing, which uses transparency effect for each pixel coloring forces the accelerator to read the color value stored in the frame buffer first, and then, after the texture is laid, to write the new value back into the buffer. GeForce3 Ti500/200, which is equipped with a more efficient "split" memory controller, feel much more at home in this case than RADEON 8500.

With multi-texturing enabled RADEON 8500 draws 6 textures in each layer at a time, which makes the total number of requests to the buffer 6 times smaller, correspondingly. So, it safely overtakes GeForce3 Ti500/200, which can lay the maximum of 4 textures in a pass.

There is a good test to measure the effectiveness of HSR algorithms, i.e. HyperZ II by RADEON 8500 and Z Occlusion Culling by GeForce3 Ti500/200. This test is called VillageMark and measures the performance of the graphics accelerators when a scene with high Overdraw is displayed:

At low resolutions HyperZ II shows almost twice as high efficiency compared with HSR implementation by GeForce3, but at higher resolutions the advantage of RADEON 8500 gets down to null little by little.

So, in synthetic tests RADEON 8500 appeared a real success, having outperformed its main competitor, GeForce3 Ti500, almost everywhere. Let us see now what is going on in gaming tests.

3D Performance: Gaming Benchmarks

Having started with synthetic 3DMark 2001 benchmarks in the previous section, we continue our review with the gaming tests:


The developers have added a bit of randomness to this test, so the results always present a sheer brain-cracker. CPU is loaded very heavily here, which you can see on the diagrams: the results hardly depend on the screen resolution. Therefore, if RADEON 8500 failed in this benchmark, we wouldn't take it very much to heart. But the new offspring of ATI even beats GeForce3 Ti500 a little in low-detail mode and runs neck and neck with it in the high-detail mode.


This gaming test doesn't only boast much more textures involved than the previous test, but also loads the T&L unit and vertex shaders considerably greater. The number of polygons in the scene has become significantly higher, and the entire model animation was set via vertex shaders. That is why RADEON 8500, which showed excellent performance in synthetic tests of T&L and vertex shaders performance, turned out a real leader here. An extra factor that told on this tremendous boost of RADEON 8500 performance was high Overdraw rate in this benchmark, which offered HyperZ II of RADEON 8500 a great opportunity to show its advantages.

An unpleasant fact is that although RADEON 8500 appeared a leader in this test, its performance dropped much greater when switching from 16bit to 32bit than by GeForce 3 Ti500/200.


Here, in the simplest benchmark, RADEON 8500 shows about the same results as GeForce3 Ti500. And again we notice a vast performance plunge when switching over from 16bit to 32bit mode.

Coming closer to GeForce3 Ti500 in 16bit modes, RADEON 8500 falls back seriously in 32bit mode. At first glance, this signals that pixel shaders are processed too slowly or the driver is ill-optimized, but having watched the test carefully we discovered that the minimum FPS rates were achieved for scenes with grass and trees, and the maximum FPS rates - for scenes with the lake surface covering most of the screen. If we take into account that grass and leaves on the trees consist of polygons with transparent textures, we will right away recall the unpleasant results of the synthetic fillrate benchmark.

But drawing the lake surface, which is created with pixel shaders, arouses no difficulties for RADEON 8500.

One way or another, this test unfortunately leaves GeForce3 Ti500 unattainably far ahead.

The Quake3 Arena testing was performed with maximum quality settings. Tri-linear filtering and texture compression were enabled.

The sensational "optimization" of the 4.13.7206 driver, which worsened the image quality in Quake3 Arena really significantly, is now gone. Now the pictures obtained by RADEON 8500 and GeForce3 Ti500/200 are of about the same quality (however, this review devotes a whole section to the image quality discussion, so the exact meaning of this "about" still has to be figured out). RADEON 8500, however, doesn't manage to beat GeForce3 Ti500: it falls some 2%-5% behind and this gap grows bigger as the resolution goes up.

Again, we witness a noticeable performance drop by RADEON 8500 when we switch over from 16bit to 32bit mode.

In this game RADEON 8500 appears in an even worse situation: its lag behind GeForce3 Ti500 grows up to 10%. Looking back at the perfect results of RADEON 8500 in Direct3D, we could make the conclusion that there's a good field of work for ATI software developers: OpenGL driver optimization.

So, the results of gaming tests let us claim that the average performance of RADEON 8500 is almost the same as that of GeForce3 Ti500, and better performance in Direct3D is "compensated" with a little lag in OpenGL.

3D Image Quality: Texture Filtering and Full-Scene Anti-Aliasing

To begin with, we would like to say that we will compare the quality of tri-linear and anisotropic filtering by RADEON 8500 and GeForce3 Ti500. For our investigation we took a scene from Quake 3 Arena:

Tri-linear filtering:

ATI RADEON 8500
  NVIDIA GeForce3 Ti500
 

With MIP-Levels
ATI RADEON 8500
  NVIDIA GeForce3 Ti500
 

As it can be seen from these screenshots, RADEON 8500 uses approximation for some reason instead of tri-linear filtering. The same situation occurs in Serious Sam and American McGee's Alice. This makes the picture quality suffer, sometimes it suffers really badly. Of course, this is a shortcoming: why should they use approximation if the performance drop even with the enabled tri-linear filtering would be very small? Probably, only ATI engineers would be able to answer this question.

Anisotropic filtering:

ATI RADEON 8500
  NVIDIA GeForce3 Ti500
 

The quality of anisotropic filtering by RADEON 8500 and GeForce3 turned out to be almost equal. However, RADEON 8500 has some artifacts near MIP-levels borders because it doesn't allow enabling tri-linear and anisotropic filtering together.

However, here RADEON 8500 shows something to be proud of and we can't help showing it to you:


The performance drops so little when anisotropic filtering by RADEON 8500 is enabled that you can surely set the maximum level of anisotropic filtering right away without any hesitation and forget about it. The same trick couldn't be applied to GeForce3 Ti500 though: the performance drop may reach 50% in case of max. anisotropic filtering settings.

And now the most interesting thing comes. Let's take a glance at SMOOTHVISION, a new full-scene anti-aliasing technology in comparison with the full-scene anti-aliasing with multisampling from NVIDIA. The driver allows enabling SMOOTHVISION in Direct3D as well as in OpenGL, and the image quality by RADEON 8500 is the same in both cases.

So, let's compare the screenshots taken from Homeworld Cataclysm for RADEON 8500 with SMOOTHVISION ("Quality" mode) with the screenshots taken for GeForce3 Ti500:

No FSAA

FSAA 2x on NVIDIA GeForce3 Ti500

FSAA 4x on NVIDIA GeForce3 Ti500

Quincunx on NVIDIA GeForce3 Ti500

SMOOTHVISION 2x Quality on ATI RADEON 8500

SMOOTHVISION 3x Quality on ATI RADEON 8500

SMOOTHVISION 4x Quality on ATI RADEON 8500

SMOOTHVISION 5x Quality on ATI RADEON 8500

SMOOTHVISION 6x Quality on ATI RADEON 8500

In the Performance mode we can see a slight change in the quality of polygon edges smoothing, which is mostly noticeable with SMOOTHVISION 2x and almost invisible with SMOOTHVISION 6x. And the second observation is that the text appears blurred like by NVIDIA GeForce2 or Vodoo5 5500.


SMOOTHVISION 2x Performance on ATI RADEON 8500

SMOOTHVISION 6x Performance on ATI RADEON 8500

Thus, we can easily state that the quality of SMOOTHVISION from ATI is much better than the quality of full-scene anti-aliasing from NVIDIA. Note that even 3x anti-aliasing performed by RADEON 8500 is no worse than 4x method performed by NVIDIA, not to mention those modes of SMOOTHVISION that involve more samples.

One more pleasant feature of SMOOTHVISION is that now there are almost no "dead angles", i.e. anti-aliasing is good at nearly any angle of slope.

Unfortunately, in spite of all its advantages, SMOOTHVISION is not yet completely debugged. For example, in Homeworld Cataclysm running in Direct3D the menu is awfully blurred, moreover depending on the selected SMOOTHVISION quality mode the blurring also differs:

Judging by the fact that the same effect is absent in case of OpenGL, we dare suppose that this is a purely software bug, which is most likely to be corrected in one of the upcoming driver updates

Now let's compare the performance drop caused by enabling SMOOTHVISION in Quake3 Arena with the drop by GeForce3 Ti500:

Unfortunately, this cool image quality shown by RADEON 8500 is far from being free of charge. However, SMOOTHVISION 4x provides much more profound anti-aliasing at almost the same expense. Hopefully, new drivers versions will reduce these performance losses.

2D Image Quality, TV-Out, DVD Playback

As usual, new ATI product boasts high image quality: even in 1600x1200 mode we noticed absolutely no blurring.

At the same time, not everything is alright with the VIVO functions of RADEON 8500. Let us try to explain it. Although Rage Theater chip supports Video-In and Out, RADEON 8500 is equipped only with a Video-Out. Moreover, some components responsible for Video-In implementation are simply absent on the PCB. We hope that in the nearest future VIVO modifications of RADEON 8500 using all the potential of the onboard Rage Theater chip will arrive.

As for the TV-Out quality, it turned out to be traditional high, as we had expected. But this is no news for you: it was quite predictable. :-)

When we tested DVD and VideoCD playback, i.e. tasks using overlays, RADEON 8500 performed nearly impeccable.

There is one fact worth mentioning separately: RADEON supports hardware MPEG2 decoding for DVD playback - Inverted Discrete Cosine Transform, or iDCT. But as we found out, not all the programs for DVD playback are able to use iDCT with RADEON 8500. For instance, the highly popular CyberLink PoweDVD 3.0 player that is shipped with Leadtek cards on NVIDIA Ti500/200 chips does not reveal any significant difference in CPU utilization while playing DVD on RADEON 8500 and GeForce3 Ti500. And the authentic ATI DVD Player that is shipped on the same CD with RADEON 8500 drivers seems to make full use of the hardware iDCT support. See for yourself: the diagram that follows below shows an average CPU utilization percentage during playing the movie called "What Dreams May Come" (PAL, 625/50 (720x576), 25 FPS, 6.14Mbps) in full-screen mode at 1600x1200x32 resolution:

The difference could be seen with the naked eye. The only question that arises is why we need hardware support for iDCT, if only 20% of the modern CPUs get utilized even without this support, and buying RADEON 8500 for a system with a weak CPU doesn't make any sense?

Concerning the quality of DVD playback, we should admit that it's pretty hard for use to evaluate it. We could only say that even after a considerably long and thorough study of the screenshot close-ups, we didn't discover any differences between the shots made on RADEON 8500 and those made on GeForce3 Ti500. We could only point out that the general brightness and color saturation differed a little, but this can always be tuned up to your taste.

Now we'll present you some screenshots from the movies:

What Dreams May Come


Shanghai Noon


The Fifth Element

Overclocking

When we overclocked ATI RADEON 8500 graphics card, it worked well at 305MHz core and 295MHz (590MHz DDR) memory frequencies without any extra cooling. This is a comparatively low result, taking into account that the nominal clock frequencies for RADEON 8500 are 275MHz/275MHz (550MHz). Higher overclocking would require extra cooling solutions, and we'll probably discuss this aspect in our upcoming article.

Conclusion

So, the worthy competitor for NVIDIA has eventually appeared. ATI RADEON 8500 graphics cards don't yield in performance to the cards based on NVIDIA GeForce3 Ti500. At that, RADEON 8500 as well as GeForce3 not only conforms with DirectX 8.1 specification, but also sports new 3D technologies improving image quality noticeably. These technologies are, for example, TRUFORM and SMOOTHVISION, and their support is not provided by GeForce3 Ti500/200.

RADEON 8500 boasts excellent 2D image quality and TV-Out implementation, which has always been a strength of all ATI products.

The quality of drivers always matters a lot: surprising as it may sound, ATI software-developers at last released high-quality drivers almost together with the first retail graphics cards. Although there are still some things to work on in the drivers, the situation is incomparably better than at the times when RADEON 256 was launched.

And the last thing we have to highlight here is the price, of course. Now the prices for the ATI RADEON 8500 boards are comparable to the prices of the GeForce3 Ti200-like graphics solutions. We mean, now ATI RADEON 8500 cards boast the best price-to-performance ratio. Note that their performance is very close to that of GeForce3 Ti500, which has been the fastest graphics solution lately.

Nevertheless, we should note that the production cost of ATI RADEON 8500 boards is hardly lower than the one of boards based on NVIDIA chips, therefore NVIDIA and the manufacturers making solutions on its chips have a vast field for price cutting. If this happens, ATI RADEON 8500 will no longer be in a favorable situation. But we, the customers, will anyway love to see the cut-throat competition causing a price drop, of course.

So, if you are an ATI worshipper, go for ATI RADEON 8500 and the card won't disappoint you, believe us.

If you prefer NVIDIA products, then just wait for GeForce3 Ti500/200 graphics cards to get cheaper, because the arrival of ATI RADEON 8500 made the position of GeForce3 Ti500/200 cards very unstable with their current pricing.

And if you haven't yet made up your mind, we hope that our review will help you with that, and the summary of ATI RADEON 8500's advantages and disadvantages will allow you to consider all the pros and cons.

Highs:

  • Highest performance in Direct3D, excellent performance in OpenGL;
  • Complete support for DirectX 8.1;
  • TRUFORM technology;
  • Best quality FSAA;
  • Low performance drops caused by anisotropic filtering;
  • Excellent image quality;
  • Multi-monitor configurations support;
  • High quality TV-Out implementation;
  • DVI support;
  • High quality mounting.

Lows:

  • Tangible performance losses with enabled SMOOTHVISION;
  • Large performance drops when switching from 16bit to 32bit modes (unlike NVIDIA GeForce3);
  • Approximation instead of tri-linear filtering in OpenGL and Direct3D;
  • No support for simultaneous usage of anisotropic and tri-linear filtering;
  • Some bugs in the drivers;
  • No Video-In;
  • Low overclocking potential;
  • Not very rich set of supplied accessories.

 
Pages: [ 1 | 2 ]

<%BANNER[banner_468x60_f]%>

Discussion

Comments currently: 1
Discussion started: 01/30/06 01:42:16 AM
Latest comment: 01/30/06 01:42:16 AM

View comments

You must log in to add comments.

Forgot password? Registration

remember me



Latest materials in Video section

Article Rating

Article Rating: 10 out of 10
 
Rate this article:
Excellent
Average
Poor