by FastSite
05/07/2001 | 12:00 AM
We have devoted quite a lot of time to testing graphics accelerators of all kinds in gaming applications. However, games are far not the only field where the today's graphics cards can show their best. The more powerful get the CPUs and graphics processors, the more efficient can be a PC used for 3D graphics development and processing. There are quite many well-known programs dealing with 3D modeling, however, one of the most widely spread applications appears 3D Studio MAX. This is one of the reasons why we decided to write an article where we consider the performance of all the today's graphics cards in this application. For our tests we took the following accelerators: Creative 3D Blaster Annihilator 2 (GeForce2 GTS), SUMA Platinum GeForce2 MX, ATI RADEON VIVO, Matrox Millennium G450 and 3dfx Voodoo5 5500. Although this article may seem pretty specific, it is aimed not only at those of you who are experienced in working with 3D Studio MAX, but at a much broader audience. Particularly for those guys who are just considering the possibility of doing some 3D design on their PCs, we included the basics of 3DMAX, which will be also very helpful for proper understanding of the obtained results. Therefore, we would strongly recommend reading the info that precedes the benchmarks area, since there you may find answers to many questions, which may turn up after checking the diagrams.
Using a 3D modeling software is very similar to filming with a video camera a certain area full of objects you created. Every project created in 3D modeling applications and in 3DMAX in particular, usually starts with building the basic geometry and introducing all the necessary modifications to it. All the work connected with the objects creation, positioning and modifying is basically carried out in the so called Viewports:
3DMAX allows modeling almost everything you wish by means of such basic objects as cubes, spheres, cylinders and cones and their modifications:
For example, the geometry of a photo camera consists only of modified basic objects:
As soon as the basic objects are created and placed accordingly, the "materials" can be applied to them, i.e. the textures can be laid over the geometry. As you may notice the textures shown in viewports are quite schematic. In fact, the displayed image quality here was optimized in order to speed up the whole thing:
Actually, the textures in 3DMAX viewports aren't completely loaded to the graphics memory, but are displayed as miniatures with adjustable size from 128x128 up to 1024x1024 pixels.
And only after the scene is completely designed, the CPU visualizes it by final software rendering. Note that in this case the image quality of the final software rendering is much higher and more precise than that the graphics card provides in viewports. And the explanation is quite simple: none of the today's graphics accelerators supports as many visual effects as necessary retaining the high fps rate at the same time. Therefore, we have to put up with very slow but very efficient software rendering. Of course, the fps rate when rendering the finalized image is very small, some frames of complex scenes may be rendered for hours. However, in this case the fastness is not the most important thing to hunt for, unlike the work in viewports, for instance. During the final rendering of the scene image quality is much more important. Especially since you needn't control this process and can leave your PC rendering the modeled final project during the night time, for example.
Definitely, you should understand very well that you work on your 3D project in the viewports and the fps rate provided by the graphics card determines if you will feel comfortable when working or not and doesn't tell on the results of the final rendering, carried out by the CPU after the scene construction is completely over.
When working in 3DMAX, we expect our graphics card to process complex geometry quickly enough, since the amount of polygons in a project of medium complexity is exponentially higher than that in a medium Quake3 gaming level.
In other words, when the scene is displayed in the viewports, the unit responsible for geometry processing bears the major workload and it is this particular unit as well as the CPU, which determine the graphics card fastness. As we have already mentioned above, all the textures, lighting and effects are applied by the CPU only after the project is finished that is why we won't consider this stage in our current article.
In games the computing capacity of the graphics accelerator is involved in processing larger textures and applying various effects to these textures. As for the amount of polygons in the today's games, they aren't that numerous at all. That is why the number of rendering pipelines and texturing units matters so greatly for games.
So, all this means that in 3DMAX the graphics card is loaded in an absolutely different way as in games. In other words, in games the stress in made on the image quality while in 3DMAX - on how fast the viewports are refreshed.
On the one hand, 3DMAX requires extremely fast processing of complex geometry in viewports. But on the other hand, when the project is nearly finished the designer can't overcome the temptation to take a look at the entire thing not only in the wireframe mode, but with all the textures and lighting applied, which may drop down the fps rate significantly. That is why in order to allow balancing between the quality and performance, the developers of 3DMAX application included several rendering methods for viewports. Since this article is devoted to 3DMAX, we decided to mention only the basic viewports rendering methods starting from the one providing the lowest quality and finishing with the one offering the best image quality. Getting a good idea of the major principles of these rendering methods will be very helpful for proper understanding of further tests.
This is the fastest method but at the same time it provides the smallest level of detail and hence the lowest image quality.
In fact, it would be incorrect to speak about any image quality here, because all the objects are represented only as rectangles of the corresponding size. This method will be suitable only for viewing very large scenes if the graphics card in your system is too slow. Of course, you won't be able to edit the objects in this case.
The wireframe method allows displaying all objects as wire-models with only ribs of the polygons seen.
This method is very often used for geometry modification since it allows you to see all the "insights" and "reverse sides" of the objects.
With this method used only polygons are displayed in the viewports.
This method is a sort of an intermediate stage between Wireframe and Smooth and offers a certain compromise of performance and quality.
This method allows not only removing sharp edges of the objects displayed but also applying lighting.
This method is very often used for texturing and further editing of the laid textures.
Note that 3DMAX allows setting different rendering methods for different viewports simultaneously.
In order to illustrate the grave difference between the best rendering quality provided by the graphics card from the final rendering provided by the CPU we would like to offer you a final screenshot of the scene considered above:
This is very good example proving once again that the image quality in viewports is sacrificed for the sake of fast operation.
In Wireframe mode 3DMAX allows using anti-aliasing. This feature can be activated only in case your graphics card runs with an OpenGL driver. Enabling Anti-Aliasing inevitably leads to the lines getting somewhat thicker, which makes the impression of lower resolution set.
As we saw, with all other rendering methods used the image quality in viewports provided by different graphics cards was almost identical. However, when rendering with the enabled anti-aliasing the image quality in systems with different graphics cards varied quite significantly.
As you can see yourself, anti-aliasing on Voodoo5 makes it impossible to work with the objects because the lines become incredibly thick. Thicker lines overlay one another making the viewport in wireframe mode a total mess. Therefore, we didn't test Voodoo5 with anti-aliasing enabled.
3DMAX offers no full-scene anti-aliasing, because no matter what graphics card you have, the performance will drop too greatly, which is unacceptable for efficient work in viewports as we have already said. Even the "Enable FSAA" option in Voodoo5 and GeForce drivers doesn't have any effect on the 3DMAX viewports.
Dear guys, please don't be shocked to see the 5-10 fps values obtained in some benchmarks. 3DMAX doesn't have the "deathmatch" mode and the work in 3D modeling applications is totally different from playing 3D games. The designer doesn't need to be as quick and accurate as a quaker, because the modeled objects won't run away. However, the major difference between the "fps in 3DMAX" and "fps in Quake3" is the fact that the 3D games hero is usually moving with some constant speed while the objects of 3D scene in 3DMAX move as fast as the designer wants them to. In other words, Quake3 is a real-time game, while the speed of objects in 3DMAX viewports is set by the designer. Therefore, in case the graphics card lacks computing capacity, the gamer will have to put up with some frames falling out while the 3DMAX designer will be able to adjust the objects speed so that to get some acceptable fps rate. The inverse relation between the fps and the moving speed can be clearly seen in the following case. Set the resolution and the level of detail in Quake3 to the maximum and run. And then walk along the same way. You will notice that in the second case there are much fewer missed frames than in the first one. And if the games usually offer two different moving speeds, such as walking and running, then the 3D modeling applications allow changing the speed of the moving objects in much greater intervals. Moreover, in order to increase the performance in 3DMAX you can disable all the objects, which aren't used at the moment, i.e. you can select only those objects, which you need to be displayed in the viewports at a certain moment of time.
So, if you see that some graphics cards showed the result of 5fps, it doesn't at all mean that normal work is no longer possible. It simply shows that the work is slowed down. In fact, you can work in 3DMAX even with the fps rate at 0.5, however, in this case everything will be so awfully slow that you'd better forget about efficiency. The ideal fps rate is just the same as in most games, namely 50-100, however, it can be obtained only for very simple scenes.
The today's graphics market offers two types of graphics accelerators: the so called gaming and professional ones. According to a widely spread opinion, professional graphics cards are absolutely useless for gaming needs and very efficient for all sorts of 3D modeling applications, while the gaming cards can suit only for playing games. We would like to stress right away that in this article we do not intend to argue about the abilities of professional graphics accelerators in other 3D modeling applications. We would only like to disagree with the general statement about professional graphics cards being much more powerful than the gaming ones or being a must for those who want to work in 3DMAX.
A few months ago we got hold of a 700-dollar Oxygen GVX1, which turned out slower in 3DMAX than NVIDIA GeForce256, despite an external geometric coprocessor (and the corresponding support in the drivers). It is exactly because of the lower performance compared to the gaming accelerators that we didn't include any professional graphics cards in our today's tests.
We would also advise you to check out the graphics cards tests carried out by the 3DMAX developer company here.
At first, we would like to explain why we tested not with the newest 3DMAX version 4, but with the version 3.1. The thing is that version 3.1 is the most widely spread 3DMAX version today and hence there are quite many additional plug-ins written for it. Besides, we think that all the tests should be carried out in "real conditions ", i.e. with the software currently used by most people, and not with the software, which will become popular some time in the future. Of course, we tried a couple of benchmarks on version 4 as well, and we can say that it is not slower and sometimes even faster than the third version. However, we are going to offer you the article devoted to graphics cards tested in 3DMAX version 4 somewhat later anyway.
Also we decided to test the cards under WindowsNT 4.0, since all OpenGL drivers are optimized for this particular OS. In Windows 98 and Windows 2000 OpenGL drivers are somewhat slower than in NT and even the possibility to work via Direct3D doesn't save the situation, since Direct3D drivers by ATI and NVIDIA display transparent objects incorrectly and the driver by Matrox is too slow to compete with all the others.
When we worked on this article we couldn't help taking a closer look at a fresh new graphics cards benchmark for 3DMAX 3.1 by SPEC. This benchmark is based on a script including a set of scenes for 3DMAX 3.1. However, this benchmark turned out extremely sensitive to all sorts of factors and we failed to get any results on any other graphics cards besides NVIDIA's ones because of various breakdowns: starting from the script crashing and finishing with the PC restarting. Moreover, even when testing NVIDIA graphics cards, the benchmark showed incorrect performance results. For instance, the performance of GeForce2 GTS was only twice as high as that of TNT2. At last, the performance turned out very much dependent on the HDD speed even in case of 512MB RAM available. Having got tired of all these problems, we made up our mind to give up this benchmark and to use the good old method suggested by 3DMAX developer company.
The second distributive CD of 3DMAX 3.1 contains a folder called "Benchmark", which includes some scenes recommended by the application developer for testing needs. Each file is aimed at checking the implementation of some selected feature in the graphics card driver as well as in the hardware. However, the distributive doesn't include any tests, which could demonstrate the overall performance of this or that graphics accelerator and not just the fastness of some implemented functions. That is why we added one more scene full of transparent objects, light sources and textures. Further on we will comment on the peculiarities of each particular benchmark, so that you didn't get lost. Well, let's go!
In our investigation we included the following graphics cards:
The testbed was configured as follows:
The software used included:
1. The first benchmark is a certain "stress-test". It plays an animated scene in four viewports at the same time. However, the rendering methods set in these viewports are different. The scene in two upper viewports is in "Wireframe" mode, in the lower left viewport - in "Smooth + HighLights" + "Edged Faces", and in lower right viewport - in "Smooth + HighLights".
Also we enabled Anti-Aliased Lines for viewports working in Wireframe mode. Since only upper viewports worked in this mode, the thing didn't influence the lower viewports at all.
This scene contains very few polygons, only 24 thousand. However, since the animation was played in all four viewports simultaneously, the resulting fps rate appeared not very high. The performance drop with Anti-Aliasing enabled was very insignificant for GeForce2 GTS/MX and RADEON, while by G450 it was just dramatic. As you will see later on, even in case of impeccable anti-aliasing quality, it is absolutely unacceptable for G450, because it causes exponential performance drop.
2. The second benchmark represents a scene with 7 standard primitives, which make the scene complexity equal to 10 thousand polygons.
There are six static objects in the scene and the seventh one is moving slowly across the entire scene passing through other objects. This benchmark checks if the intersecting objects are displayed correctly and if the graphics card and the driver cope with this task fast enough.
Voodoo5 appears twice as slow asGeForce2 and 1.5 times slower than RADEON and G450. Since this scene is very simple, then the problem must lie with the driver, for sure. It looks very much as if in every frame the drivers were drawing anew all the objects instead of redrawing only those, which have been changed. The software rendering of the scene appeared impressively fast, having turned out much faster than the intermediate rendering carried out by the graphics cards. We think it was possible due to light geometry used (only 10 thousand polygons) and as is known, the driver used for the software rendering is very well optimized for light geometry.
3. The scene from the third benchmark shows a ball, which is moving very slowly against the background geometry made of 15000 polygons.
The ball doesn't cross any other objects. Since it moves very slowly, then the "perfect" driver will make very few changes to each further frame. In other words, this benchmarks checks if the graphics card is capable of drawing anew only those objects, which got really changed.
Just like in the previous benchmark, the software rendering proved highly efficient. No wonder, since 3DMAX developers work on its optimization from the very first version of the application. As for the graphics cards, the situation repeats. Voodoo5 suffers some noticeable problems with the drivers, while Matrox G450 turns out unexpectedly good.
4. This benchmark shows if the graphics card is good at processing very complex geometry (imitating mountains). The overall amount of polygons used here makes 200270.
The leading positions were won by NVIDIA based graphics cards, as we have expected actually. Then to our great surprise comes Matrox G450. Well, Matrox' software developers seem to have set down to work, really.
5. The fifth benchmark is devoted to testing the graphics cards abilities in terms of processing complex geometry only. This time the amount of polygons nearly doubled and made 376 thousand. The same surface, as in the previous benchmark scene, is now covered with buildings.
This benchmark can easily bring any graphics card to its knees: the average fps rate hardly reaches 3 frames. However, you should bear in mind that this is just a benchmark testing the card's geometric performance. The file was surely created not with the 3fps. Each building was designed in a separate file and when it was added to the entire scene, some part of the geometry was disabled in order to increase the performance.
6. Having taken a look at the graphics cards performance when working with geometry, we suggest passing over to the imitation of multiple light sources. This test offers 8 SpotLights, which are constantly moving and lighting some kind of asteroid.
We should point out that imitating the effect made by SpotLights is a much more resource-hungry process than the imitation of Omni or Directional lighting.
Again NVIDIA graphics cards leave all their competitors very far behind. And again Voodoo5 drivers make this card lag in the very end. However, Matrox drivers also aren't optimized for this kind of work.
7. Here we see the same asteroid, but this time it is lit by 8 Directional lights. Directional lights are slower than Omni lights, but faster than SpotLights.
Well, what do we see now? NVIDIA and ATI graphics cards are running as fast as they were in the previous test. The speed of Voodoo5 doubled and that of G450 tripled, proving once again that Directional lights are faster than Spotlights (just what we have said).
8. One more time we see the same asteroid with the lights coming from 8 different sources. But now these are all Omni lights, the fastest lights in 3DMAX.
Well, NVIDIA graphics cards seem not to care at all what type of lighting is currently used in the scene. In all three benchmarks GeForce2 based graphics cards perform on the same level, which means that the drivers are well optimized. In fact, we would be pretty surprised to find out that the drivers, which are updated every now and then were optimized badly. By the way, the same thing is valid for RADEON as well: in all the three lighting benchmarks the cards demonstrated the same performance level, although we can't say that its drivers are also fully optimized for this type of effects, because the card is almost two times slower than NVIDIA's offspring. In other words, the driver is well optimized for all types of lighting, however the optimization is still insufficient for the card to catch up with the products based on NVIDIA chips. As for Voodoo5 and G450, they turned out slower than in the previous test, showing that the driver developers didn't pay enough attention to Omni lights.
9. The ninth benchmark is a scene with light geometry, including only 4500 polygons, which occupies the entire viewport. It is aimed at testing the rasterizing speed.
When the camera is moving, the graphics card should rasterize big and small polygons (relative to the screen size).
Just as in the previously described benchmarks, the graphics cards based on NVIDIA chips appear in the leading positions, while RADEON is just a little bit behind them. Voodoo5 is the last, as usual.
10. The next benchmark is devoted to work with textures. The file contains a lot of textures and very little geometry (224 polygons). As for the benchmark, it is just a rotating spherical polygon with the facets covered with 48 different textures.
Very little geometry and many textures involved give us a perfect idea of how fast the graphics cards can process these textures.
Cool! This is the first benchmark where ATI RADEON appeared a bit ahead of NVIDIA GeForce2 . Voodoo5 turned out a complete failure, since it didn't manage to display a single texture.
11. This is a fully textured room with a camera moving inside. The file includes 12 thousand polygons and the geometry quality is very close to that typical of real projects. Also there are 7 light sources inside the room.
Since there are multiple light sources and quite complex geometry, this benchmark can be regarded as close to "general" scenes. GeForce2 again leads the procession.
12. Animated "waves" with the 114KB texture laid over them show how fast the card can deform very light geometry and modify smaller textures.
And the geometry in this scene is really extremely light: only 838 polygons.
This is the second purely texturing benchmark with light geometry and minimal lighting. The first position still belongs to RADEON, which means that its driver is well optimized for textures processing. Voodoo5 appeared surprisingly fast here. However, this is a really easy benchmark. The difficult one turned out too tough for Voodoo5.
13. This benchmark runs with different speeds in the Wireframe mode. 111 thousand polygons in Wireframe mode will be a really tough test for any modern graphics card. Just as in the very first benchmark we enabled Anti-Aliasing here:
Here the winner turned out RADEON, which showed that you needn't sacrifice fastness for the sake of good Anti-Aliasing effect. Then almost level with RADEON come NVIDIA based graphics cards. The performance dropped just a tiny bit with Anti-Aliasing enabled. Matrox again surprised us with high results, which fell down unpleasantly greatly as soon as we enabled Anti-Aliasing. Voodoo5 lags behind everyone else and its Anti-Aliasing mode is simply unbearable.
All the benchmarks described above are recommended by 3DMAX developers. However, as we have already seen, they are aimed at testing different functions and their implementation separately from one another. Since there are no "general" tests there, we decided to add one more benchmark to this set: a scene with 8 light sources, 61371 polygons and a great lot of transparent surfaces. The file with all textures makes 6MB total size and its complexity is quite typical of the today's 3D projects. We also included some animation to provide more realistic testing conditions: the camera is moving around the room capturing all objects. This is how the first frame looks after the final rendering is done:
Now let's see how the graphics cards will cope with this task in Wireframe mode and in WireFrame + Anti-Aliasing:
Since this benchmark is a kind of final one for the graphics cards showing their performance in Wireframe mode, we decided to base our conclusions on the results obtained in it. What do we see here? GeForce2 based cards won the lead. GeForce2 GTS and GeForce2 MX differ only by the graphics core frequency (25MHz). The twice as fast memory of GeForce2 GTS (333MHz against 166MHz) doesn't tell on the results obtained in Wireframe mode, since no textures are displayed here. RADEON proved very efficient in this test: it is just a little bit behind the leader. As for G450, we have already pointed out that its performance in Wireframe mode leaves much to be desired in case Anti-Aliasing is enabled. As a result, this mode appears absolutely unsuitable for Matrox G450. If Anti-Aliasing is disabled then the performance turns out quite all right for this graphics solution. Voodoo5 again failed to stand this test, especially since its Anti-Aliasing didn't work properly.
And now let's take a look at the results obtained for the same scene in Smooth + HighLights mode:
Like in the just considered Wireframe mode, GeForce2 based cards won the first prize. However, this time RADEON appeared really far behind them. G450, which was not bad at all in the previous test, now occupies the very last position. Voodoo5 again has hardly anything to be proud of, falling far behind GeForce2 and RADEON.
No doubt, the best choice you could make for work in 3DMAX is a solution based on NVIDIA GeForce2 chip.
GeForce2 based graphics cards led the show in all the tests except those dealing with texturing, where they yielded a couple of percents to ATI RADEON (see benchmarks 10 and 12). Different light sources were processed equally fast. Besides, the cards proved very "clever" when updating the viewports in the third benchmark. All this proves that the OpenGL drivers are very well optimized not only for gaming purposes. However, since in benchmarks 2 and 3 software rendering appeared the fastest, NVIDIA software developers will have a lot of room for further driver improvement, that's for sure. In the meanwhile, the drivers seem to be more optimized for 3DMAX.
The tests carried out for GeForce2 GTS and GeForce2 MX based graphics accelerators showed that it is only the core frequency that influences the performance significantly. GeForce2 MX based card was constantly falling only 10% behind its elder brother. Note that the difference in core frequency between these chips makes about 10% as well: 175MHz by GeForce2 MX and 200MHz by GeForce2 GTS. All the other things, such as the graphics memory working frequency, which is so vitally important in games, as well as the number of pipelines, do not provide any advantages for work in 3DMAX. So, from the price-to-performance point of view, low-cost GeForce2 MX appears the best choice.
ATI RADEON won the second prize, which is a really good achievement, since ATI products have been always famous for the absence of well-done OpenGL drivers. When the graphics card was just announced and started selling, the drivers were so poorly optimized for 3D modeling applications that almost every time you tried to load a certain scene from 3DMAX, you got an empty viewport, i.e. the driver couldn't display anything at all. However, ATI officials kept saying that the company was working on the driver optimization not only for games but also for professional 3D applications. Well, as we see, ATI kept the promise.
Our tests showed that the drivers are perfectly optimized for textures processing and modification (benchmarks 10 and 12), that the performance in Wireframe mode is high enough (benchmarks 1, 13, 14) and that the performance drops the least of all in case we enable Anti-Aliasing (benchmarks 1, 13, 14). Also we should point out that the performance didn't change when we shifted to more complex lighting effects (benchmarks 7, 8, 9). However, there is still a fly in the ointment: though RADEON is nearly impeccable in Wireframe mode, in the second basic mode, Smooth + HighLights the results drop down nearly twice (benchmarks 11, 15). This performance lowering makes RADEON unable to compete with GeForce2. However, as we see, the company keeps working on the drivers that is why ATI may manage to develop a super-optimized driver by the time we start testing graphics accelerators in 3DMAX 4.0.
Matrox Millennium G450 appeared very hard to evaluate.
As a rule, Matrox didn't care about the implementation of those functions, which were necessary not only for work in 3D modeling applications but also for proper displaying of the separate scenes. The performance of G450 in Wireframe mode (benchmarks 13, 14) corresponds to the cost of this graphics solution, however as soon as Anti-Aliasing is enabled, you'd better forget about any efficient work: the performance drops down about 10 times. And even though theoretically you can do without Anti-Aliasing, low performance in Smooth + HighLights mode (benchmark 14) make further use of G450 in 3DMAX absolutely impossible.
Voodoo5 appeared a total disappointment.
It turned out a complete failure in all the tests, which proves that its drivers aren't at all optimized for professional usage. The last straw that broke the camel's back was the Wireframe mode with Anti-Aliasing enabled. All in all, Voodoo5 is the worst thing a 3D graphics developer could ever think of.