by Michail Djagterev
03/22/2003 | 01:10 PM
February 2003, people at Futuremark, previously known as MadOnion.com (and before that – the same as now – Futuremark), released their 3DMark03 benchmarking package intended for measuring speed characteristics of modern PCs in 3D games. Right after its appearance, 3DMark03 provoked a lot of critical remarks from both: graphics chip makers (NVIDIA was among the most disappointed) and graphics cards reviewers. One of the main concerns was the low performance of modern graphics cards, particularly GeForce4 Ti4600 based ones, which didn’t correspond well to their pretty fast performance in present-day computer games.
<%BANNER[article]%>Anyway, 3DMark03 is a benchmarking tool that helps to check the performance of graphics cards in next generation 3D applications. It means we can have a peep into the future and get to know beforehand the behavior of existing graphics cards in games that are going to use the latest advances in computer graphics.
We tested a bunch of thirteen graphics cards available in the today’s market. So, we are sure that you will find some really useful information in our article.
We tested the following graphics cards:
So, we gathered together nearly all the graphics chip now offered by ATI and NVIDIA and added a card based on SiS’ chip. Ten of the cards have 128MB of graphics memory onboard. It’s the most optimal and usual memory amount found in high-performance gaming graphics cards. Three cards have 64MB of memory. We can single out RADEON 9500 among them as the only DirectX9-supporting card with 64MB memory. One graphics card, GeForceFX 5800 Ultra, has 128MB of memory of the new type – DDR II. So far, the future of this card is not clear, so we can’t tell whether we will see it in the open. Anyway, it’s the newest and fastest GPU from NVIDIA for today.
The cards worked with the following drivers:
You see that we used two versions of Detonator drivers. When 3DMark03 appeared, the 42.86 drivers were fresh, but soon Detonator 42.68 appeared. Some sources say the last version was optimized for 3DMark03. GeForceFX based cards do show notably higher performance with these drivers in 3DMark03, but GeForce4 Ti GPUs experience some problems. For example, there may be image artifacts like a wide horizontal band across the screen:
You can’t play like that! So we tested all GeForce4 based cards with Detonator 42.86 and GeForceFX – with Detonator 42.68.
The major goal of our testsession was to reveal the performance of the above listed graphics cards in 3DMark03 as well as the performance drop they would suffer on turning on post-processing, full-screen anti-aliasing and anisotropic filtering. We ran each gaming test in the following modes:
All the cards were tested with “default” driver settings except those cases when we had to force full-screen anti-aliasing or anisotropic filtering via the drivers. It looks like 3DMark03 can’t always determine whether the card supports FSAA or anisotropic filtering, as well as the maximum supported anisotropy level. That’s why when forcing these settings we set FSAA 4x and 8x anisotropy level. As NVIDIA GeForce4 MX440-8x allows setting only 2x anisotropy, we used this as a maximum level for this card.
All we have to do now is to test each card in each mode and compare their results. We will explain all cases when a card cannot pass a test, and also show screenshots with image artifacts indicating a failure.
So, let’s get started.
This test uses only DirectX7 instructions and it was the only one that ran on each of the thirteen cards without any problems. Post-processing is off, tri-linear filtering is on:
GeForceFX 5800 Ultra based card wins this test in all resolutions thanks to the highest frequency of the GPU – 500MHz, and the formidable peak fill-rate – up to 4000MTexels/s. RADEON 9700 PRO fell behind as its core works only at 325MHz, and the peak fill-rate with multi-texturing equals 2600MTexels/s.
GeForce4 Ti4800 differs from GeForce4 Ti4600 only in its AGP 8x support. So we can estimate the advantage of the new interface. In fact, AGP 8x brings just a small performance growth here: about 1-1.5 fps. The “Game 1 – Wings of Fury” test doesn’t have huge amounts of textures and vertex matrices: the data easily fit into 128MB of local graphics memory. That’s why there is so little benefit provided by theoretically faster AGP 8x.
As far as the amount of graphics memory is concerned, we can see by the example of the two versions of RADEON 9500 that 128MB instead of 64MB provide a nice speed-up: over 10% in high resolutions.
GeForce4 MX and Xabre 600 are outsiders in this test. In fact, “Game 1 Wings of Fury” was the only test to run on GeForce4 MX. As for Xabre 600, it works smoothly here, but shows poorest performance among all.
Now we turn post-processing on:
Xabre 600 was the only card without post-processing support. That’s why it has zero in the diagram. In comparison with the previous mode, this one brings no significant changes. Moreover, post-processing even led to some speed growth in 1600x1200 resolution. Otherwise, the situation is the same with the same leaders: GeForceFX 5800 Ultra, RADEON 9700 PRO and GeForce4 Ti4800.
Now let’s see the effect of anisotropic filtering on the cards’ performance.
All the cards suffer a significant performance drop. And once again, GeForceFX 5800 Ultra wins the test. Anisotropic filtering results in a performance drop of 1.1–1.6 times by GeForceFX 5800 Ultra. The GeForce4 Ti based cards slow down by 1.5–2 times. We also see the same 1.5 times performance reduction by RADEON 9500 and 9700.
As for Xabre 600, this graphics solution doesn’t support anisotropic filtering at all. Among surprising results, we’ve got GeForce4 MX beating GeForce3 Ti200 and GeForce4 Ti4200-8x. The point is that this GPU only supports 2x anisotropy level, while other participators perform 8x anisotropic filtering.
Next, we turned on full-screen anti-aliasing.
FSAA brings some considerable distortion in the ranks. RADEON 9700 PRO is in the lead now thanks to its broader graphics memory bus. In high resolutions it wins about 20% over GeForceFX 5800 Ultra. Overall, GeForce4 Ti cards lose to RADEON 9500 PRO and 9500 with 128MB or 64MB of memory. The only thing to comfort NVIDIA fans is the win of GeForce3 Ti200 over its immediate rival, RADEON 9100, in 1024x768 and 1280x1024 resolutions.
But there was one problem with 128MB RADEON 9500. It was the only of ATI’s graphics chips to produce the following glitch: vertexes of some polygons are processed incorrectly, so there are lines in the image that are supposed to be invisible. You can see this artifact in the following screenshot:
Now, let’s say a few words about those who didn’t finish. RADEON 9000 couldn’t work stable in any resolution, so we took it off the race. All 64MB cards couldn’t manage 1600x1200 resolution, which is quite natural. Xabre 600 only supports FSAA 2x, so it’s not quite right to compare it with others. Moreover, with FSAA turned on, Xabre 600 displays the image only on one half of the screen in 3DMark03, that is, it performs only a half of the job :). Of course, you can’t play like that, so we didn’t include the results shown by Xabre 600 into the diagram.
The next step includes FSAA with anisotropic filtering.
Although RADEON 9000 and 9500 PRO didn’t make it through the above mode, they did it now. Xabre 600 was not included, as it has no anisotropic filtering support. The simultaneous use of FSAA and anisotropic filtering resulted into the absence of actual leaders in this test session. Up to 1600x1200 resolution the laurels belonged to GeForceFX 5800 Ultra running neck and neck with RADEON 9700 PRO. Only in 1600x1200, RADEON 9700 PRO managed it ahead.
GeForce4 MX440-8x and RADEON 9000 managed all resolutions up to 1024x768 inclusive. In 1280x1024 they disabled anti-aliasing. That’s why they have zeroes in the two highest resolutions. RADEON 9100 and 64MB RADEON 9500 also dropped out of the race in 1600x1200.
We would like to single out the results shown by RADEON 9500 PRO. This graphics card keeps its third place throughout all the resolutions and leaves far behind (almost two times) the cards based on GeForce4 Ti4600 and Ti4800.
Well, the first gaming test, “Wings of Fury”, indicated what performance you can expect from your graphics card in a DirectX7 flight simulator. We can’t say this test illustrates the performance of our testing participants in all other DirectX7 games, but this is the only DirectX7 test in 3DMark03.
Now, it’s time we went over to other tests that require DirectX8 support.
The more complex test, “Battle of Proxycon”, uses pixel and vertex shaders of DirectX 8.1. GeForce4 MX440-8x leaves us here, because it doesn’t support this version of the Microsoft’s API. Well, we have just lost one player; let’s see what the others can show!
We will start out with the simplest task: tri-linear filtering without post-processing.
“Battle of Proxycon” test uses ver.1.4 pixel shaders. This allows rendering a scene within fewer passes than in case of ver.1.1 pixel shaders. 3DMark03 package doesn’t require the graphics card to support ver.1.4 pixel shaders to run “Battle of Proxycon”. If the card supports only 1.1 version, 3DMark03 will use it to render the scene, although in this case more passes will be needed.
It’s quite natural that the cards based on RADEON 9500, 9500 PRO and 9700 PRO chips are notably faster than GeForce4 Titanium based ones. Even the slowest RADEON 9500 with 64MB memory outperforms GeForce4 Ti4800 nearly in all resolutions.
The winner of “Battle of Proxycon” is GeForceFX 5800 Ultra, which performs best in all resolutions. RADEON 9700 PRO and 9500 PRO are second and third, respectively. These three graphics chips support ver.1.4 pixel shaders; the difference in their results is only due to the difference in their clock-rates and the number of working pipelines.
Surprisingly, RADEON 9000 works faster in this test than RADEON 9100. The reason again lies with the pixel shaders, namely with the way they have been implemented. RADEON 9000 graphics chip was developed long after RADEON 8500 (now renamed into RADEON 9100). And the pixel shaders units of RADEON 9000 have been improved over RADEON 9100. And as the execution speed of pixel shaders is crucial in “Battle of Proxycon”, the younger RADEON proves brisker than the oldie, although has fewer texturing modules and lower peak fill-rate. At the same time, we can’t find a reasonable explanation of the fact that GeForce3 Ti200, which doesn’t support ver.1.4 pixel shaders and works at lower frequencies than RADEON 9100 and 9000, manages to defeat both in this test.
Xabre 600 shows the lowest results. But the failure is not only about performance. This graphics chip cannot reproduce dynamic lighting in the scene. There are also problems with texture rendering. They result in artifacts like black dots and stripes:
Looks like SiS programmers are spending little time writing Xminator drivers. They have still quite a lot to do :) .
By the way, NVIDIA GeForce4 Titanium series also generate certain image artifacts in the second gaming test. They are not so evident as those noticed by Xabre 600, but are still very unpleasant. Thus, if you look at the haloes left by the shots or the flame of the jet nozzles, you will see very sharp color transitions. You will not notice anything like that in the pictures rendered by other cards.
We suppose these transitions appear because GeForce4 Titanium GPUs use texture compression in 3DMark03 or transform the textures into the 16-bit format to reach higher performance.
Now we turn on post-processing.
Post-processing in “Battle of Proxycon” results in lower performance. That is, the better quality of the picture is not free-of-charge here, as in the first gaming test. Otherwise, the situation remains the same: GeForceFX 5800 Ultra is ahead in all resolutions, followed by RADEON 9700 PRO. RADEON 9500 PRO is the third. We can now see clearly the advantage of 128MB memory over 64MB: the 128MB variant of RADEON 9500 is faster by about 50%. This gap was much smaller without post-processing.
Now, anisotropic filtering is on…
Enabled anisotropic filtering forced one more graphics card, SiS Xabre600 to quit the test, because it doesn’t support this filtering method. GeForceFX 5800 Ultra is still in the lead, closely followed by RADEON 9700 PRO. The performance of both cards dropped by about 1.5 times as we turned on anisotropic filtering.
RADEON 9000 and 9100 are faster here than GeForce3 Ti200: ATI’s graphics chips perform anisotropic filtering more effectively. But we just can’t explain why GeForce3 Ti200 outruns GeForce4 Ti4200-8x! The newer GeForce4 Ti4200-8x boasts a more powerful core, works at higher frequencies and uses AGP 8x (instead of AGP 4x used by GeForce3 Ti200). So, the better results of GeForce3 Ti200 can only be explained by the use of lower anisotropy levels instead of 8x. Otherwise, GeForce3 Ti200 just can’t be faster than GeForce4 Ti4200-8x!
“Game 2 – Battle of Proxycon” uses about 87MB of graphics memory to store textures and vertexes. Let’s try to load the graphics memory more by turning on FSAA 4x.
And once again, GeForceFX 5800 Ultra is far ahead. In 1024x768 resolution it’s 1.8 times faster than RADEON 9700 PRO! RADEON 9500 PRO keeps its third place, followed by GeForce4 Ti4800.
RADEON 9000 worked unstable with FSAA 4x in this test as well as in “Game 1 Wings of Fury”. So it has zero in the diagram. Xabre 600 once again shows the scene in one half of the screen only, so we will not take its results into account:
Funny, isn’t it? :)
Now, the full load: FSAA 4x plus anisotropic filtering.
You could have guessed: GeForceFX 5800 Ultra is on top, RADEON 9700 PRO is the second, RADEON 9500 PRO is the third. The amount of graphics memory does matter here: compare the results of 128MB and 64MB RADEON 9500. Other results need no comments of ours.
Futuremark Company claims that “Battle of Proxycon” is not enough to analyze graphics card’s performance in DirectX8 applications. The next gaming test, “Troll’s Lair”, uses the same technologies as Game 2, but somewhat differs from it.
So, at first we run the test with tri-linear filtering turned on.
Unlike the “Battle of Proxycon”, GeForceFX 5800 Ultra has smaller advantage over RADEON 9700 PRO here. They fight for every fps! :)
The graphics cards based on RADEON 9500 PRO and 9500 graphics chips feel more confident here than in the previous test. We can see it in 800x600 resolution, where the junior DirectX9-compatible card (64MB RADEON 9500) is 1.5 times faster than GeForce4 Ti4800, the most powerful GPU in the GeForce4 family. In high resolutions the gap in cards’ performances becomes smaller, but their positions in the race remain unchanged. RADEON 9000 outperforms RADEON 9100 in “Troll’s Lair”, which is not at all surprising, because the third gaming test is similar to the second test and also uses ver.1.4 pixel shaders, which are performed much faster by RADEON 9000.
Xabre 600 produces the same artifacts as in the previous gaming test here. So we don’t even take into account its results, especially since they are anyway very low. If we had a real PC game instead of “Troll’s Lair”, you wouldn’t be able to play it on Xabre 600…
We turn on post-processing in the next round of testing.
Post-processing brings no changes into the leading group: GeForceFX 5800 Ultra is the first, followed by RADEON 9700 PRO and RADEON 9500 PRO.
But now that the graphics memory workload has become considerably higher we can see how big is the performance difference between the 64MB version and 128MB version of the RADEON 9500 based graphics card. Of course, 64MB of graphics memory seems too little for this test, therefore 64MB RADEON 9500 loses not only to its 128MB mate, but also to GeForce4 Ti4600 and 4800 (in resolutions over 800x600) and also to GeForce4 Ti4200-8x (in resolutions over 1280x1024).
The influence of anisotropic filtering comes next.
The performance difference between GeForceFX 5800 Ultra and RADEON 9700 PRO remains pretty small in this test, but GeForceFX 5800 Ultra still doesn’t give up its leadership. RADEON 9500 PRO graphics card occupies the third place being just a step behind RADEON 9700 PRO.
Curiously enough, GeForce3 Ti200 catches up with GeForce4 Ti4200-8x in higher resolutions. In 1600x1200 the two cards show similar performance, notwithstanding differences in their frequencies as well as in graphics chip architecture. Of course, it doesn’t mean that GeForce3 Ti200 can match GeForce4 Ti4200-8x in modern games. There must be something wrong with anisotropic filtering implementation in Ti200 as we have already mentioned above.
Now let’s pass over to the tests with enabled anti-aliasing.
With 4x anti-aliasing on, ATI’s graphics solutions suffer a bigger performance drop than NVIDIA’s GPUs. As a result, the gap between GeForceFX 5800 Ultra and RADEON 9700 PRO becomes wider, while the one between RADEON 9500 and GeForce4 Ti4800 – narrower. As a result, the third and the fourth positions are occupied by RADEON 9500 PRO and RADEON 9500 128MB respectively, while the fifth position belongs to GeForce4 Ti4800, which defeats 64MB RADEON 9500 and catches up with 128MB RADEON 9500 in high resolutions.
Xabre 600 beats GeForce3 Ti200 in 800x600 resolution, but as this graphics chip outputs the picture only onto one half of the screen, we get every right to divide its results by two as well :).
Now, the full load: FSAA 4x and anisotropic filtering.
Overall, the results are similar to what we saw in “Battle of Proxycon” with enabled anisotropic filtering and anti-aliasing. The leaders are the same: GeForceFX 5800 Ultra is the first in all resolutions, RADEON 9700 PRO and 9500 PRO are the second and the third correspondingly. The graphics cards based on GeForce4 Titanium chips can compete here only with RADEON 9100 and RADEON 9000. And if we take into account that the performance difference between GeForce3 Ti200 and RADEON 9500 PRO is only 2-3 fps and the performance of RADEON 9500 PRO varies from 3.1 to 9.3fps in different resolutions, then the further discussion of graphics cards performance will simply lose sense. The speeds are too low to be compared.
Well, here we would like to finish with DirectX8 Game 2 and Game 3 tests and will pass to the last gaming test aka Game 4 – Mother Nature.This test requires DirectX9 support from all participating graphics cards. It means that only five cards of the thirteen are going to take part in this test. They are RADEON 9700 PRO, RADEON 9500 PRO, RADEON 9500 with 128MB and 64MB memory, and GeForceFX 5800 Ultra.
NVIDIA may be proud: its GeForceFX 5800 Ultra wins over the four RADEONs quite successfully. GeForceFX 5800 Ultra again dashes far ahead of RADEON 9700 Pro. The third, fourth and fifth positions have also been distributed quite logically: RADEON 9500 PRO, RADEON 9500 128MB and RADEON 9500 64MB.
“Game 4 – Mother Nature” test uses 54MB of graphics memory for the vertex buffer, 50MB for textures and 9MB for the index buffer. So, the scene in the fourth gaming test may take up about 113MB of graphics memory. It’s 1.7 times more than 64MB RADEON 9500 can accommodate, and this card has to load additional data from the system memory via the AGP bus.
But we have no great difference in performance of the 64MB and 128MB versions of the card here as well as in other gaming tests where no post-processing, 4x FSAA or anisotropic filtering were used. It means that the advantage of additional 64MB of graphics memory is not crucial even in complex scenes with high amounts of textures if none of those functions is turned on. And it doesn’t even matter whether the scene uses DirectX7, or DirectX8 or DirectX9.
Now we enable post-processing.
Enabling post-processing in this hardest gaming test increases the gaps between all graphics cards in all resolutions, except 1600x1200. GeForceFX 5800 Ultra is again ahead of all, and the performance difference between it and RADEON 9700 PRO is especially evident in 1600x1200. Here only 128MB graphics cards manage to work, just like in 1280x1024. And the performance difference between RADEON 9500 128MB and RADEON 9700 PRO makes only 0.6 frames per second. In other words, the performance difference between these two graphics cards is not proportional and looks very unusual against the background of the previous tests results. But where does this huge performance drop on transition from 1280x1024 to 1600x1200 come from?
On the one hand, there is higher graphics memory workload in this resolution and the graphics cards have transfer the data from the local graphics memory along the AGP 8x bus. But as we have already learned from the previous tests (and especially from Game 4 - Mother Nature without post-processing and anisotropic filtering), the graphics card doesn’t slow down too much in this case: 128MB and 64MB RADEON 9500 show similar performance. On the other hand, with enabled post-processing this performance difference gets even more noticeable and becomes especially evident in 1024x768. Analyzing the performance reduction because of the lack of local graphics memory, I dare suppose that in 1600x1200 RADEON 9500, RADEON 9500 Pro and RADEON 9700 PRO could work about 1.5-2 times slower than in 1280x1024 resolution, but this number would never be 2-3 times, by no means! The explanation that comes to mind is certain faults in ATI’s drivers, but this is only a supposition. We guess we can’t make far-going statements basing on results of a single 3DMark03 test.
Now we are passing over to the tests with enabled anisotropic filtering.
GeForceFX 5800 Ultra still holds its ground. The second position as usual belongs to RADEON 9700 PRO. The example of two RADEON 9500 versions with different amount of graphics memory onboard we can see that there is a very little gain derived from extra 64MB of graphics memory: only 0.5-1.2fps. The graphics cards anyway have to transfer the data via the AGP bus and they both suffer from the lack of local graphics memory. But 64MB is far not enough to work in 1600x1200 resolution with enabled anisotropic filtering, and this is a much graver limitation than those 0.5-1.2fps mentioned above.
The next step is enabled anti-aliasing.
64MB RADEON 9500 couldn’t chew up the test in this mode, which is not surprising at all keeping in mind the memory size required to run Game 4 – Mother Nature with 4x FSAA. Since in all the previous tests with anti-aliasing the 64MB RADEON 9500 worked OK, at least up to 1600x1200 resolution, we can conclude that 3DMark03 requires the graphics card to store not only the frame buffer in the local memory but also the vertex buffer. The fourth gaming test has a 54MB vertex buffer (compare to 6MB, 6MB and 19MB buffers by the three previous tests, respectively). So, no wonder that with enabled FSAA requiring bigger frame buffer, there is simply no room left in the 64MB graphics memory of RADEON 9500 to store all the data.
Well, it’s high time we ran the hardest test of all: Game 4 – Mother Nature with enabled FSAA 4x and anisotropic filtering at a time. Let’s see what will come out of it.
Even in this heaviest test, as usual, it is GeForceFX 5800 Ultra based graphics card that wins! RADEON 9700 PRO loses no more than 2.3fps, but never comes ahead.RADEON 9500 PRO and RADEON 9500 128MB split the third and the fourth places. The 64MB RADEON 9500 based card didn’t manage “Mother Nature” with both anti-aliasing and anisotropic filtering on.
Summing it up, we should acknowledge that “Game 4 – Mother Nature” demonstrated the superiority of graphics cards based on NVIDIA GeForceFX 5800 Ultra GPU in complex DirectX9 applications. This graphics card won all the tests. During the test session in Game 4 – Mother Nature we didn’t reveal any image artifacts by any of the 5 graphics cards tested. On the whole, “Mother Nature” is a questionable test. It uses an awfully complex scene that has more geometry in one frame than any modern game can boast. It also doesn’t use all DirectX9 instructions. So, we have just had a look at how well present-day graphics cards feel under such hard working conditions of tomorrow. Of thirteen cards that passed “Wings of Fury”, there are only five that were allowed to “Mother Nature” (the others just don’t support DirectX9). But even among these five cards, one couldn’t work with full-screen anti-aliasing (as it has only 64MB of graphics memory) and another – with anisotropic filtering. A sad picture, really.
From gaming tests we now turn to synthetic ones. They will help us to estimate the cards’ performance from the functional point of view.
CPU Test 1 and CPU Test 2 were run in 640x480 resolution to reduce the dependence of the results on the graphics card’s performance, especially on their fill-rate. The both tests use Game 1 and Game 2 gaming scenes (without pixel shaders and dynamic shadows). Futuremark Company claims that the results in the CPU tests still depend on the graphics card. We are going to check out how big this dependence is.
We see that the dependence is quite considerable. And “CPU Test 1” becomes the first and only test won by Xabre 600! We were quite surprised at that. In the first CPU test Xabre 600 is 4.7fps faster than GeForceFX 5800 Ultra. In fact, all the leaders of the gaming tests now lose their ground: GeForceFX 5800 Ultra is slower than GeForce4 Titanium based cards and even slower than GeForce3 Ti200, while RADEON 9700 PRO loses to RADEON 9500 PRO and both RADEON 9500 versions. Yeah, the graphics card affects the results of the CPU tests in 3DMark03, but the results are no proportional to the performance of the card in gaming tests. The results let me draw only one conclusion: we shouldn’t rely upon CPU benchmarks integrated into 3DMark03. They may show quite well the performance of different processors when you use them with one and the same graphics card. But when using different graphics cards, you can’t estimate CPU performance accurately enough.
Fill-rate measuring tests were run in 1024x768 with standard 3DMark03 and drivers settings.
The results of the FillRate test are somewhat confusing.
In Single-Texturing mode, when there is only one texture laid over one surface, RADEON 9700 shows the best performance. Its eight-pixel pipelines provide the maximum fill-rate of 2600MTexels/sec in this mode, while GeForceFX 5800 Ultra with only four pixel pipelines ensures 2000MTexels/sec. But they swap places in case of multitexturing. Now the higher frequency of NVIDIA GeForceFX 5800 Ultra (500MHz) allows it to outperform the 325MHz RADEON 9700 PRO by more than 1000MTexels/sec.
128MB RADEON 9500 takes the third place in the single-texturing mode. It has the same core frequency (275MHz) as RADEON 9500 PRO, but only four pixel pipelines against eight. How did it win then? It’s because the 128MB version of RADEON 9500 has a wider memory bus – 256-bit against 128-bit by RADEON 9500 PRO (and RADEON 9500 64MB). Wider memory bus determines higher performance of the 128MB RADEON 9500 over RADEON 9500 PRO and RADEON 9500 64MB.
NVIDIA GeForce4 Ti4600 and Ti4800 are following the leaders in the multi-texturing mode: they occupy the third position in the rank.
They have four pixel pipelines with two texturing units each. Together with their higher working frequencies (300MHz), this helps them to outperform RADEON 9500 PRO and other cards, but they still fail to reach RADEON 9700 PRO because of the narrower 128bit memory bus. In multi-texturing mode RADEON 9500 is slowed down by its 128bit memory bus, so that with its 4 pixel pipelines and 1 TMU per pipeline it loses even to GeForce4 Ti4200-8x and Xabre600 with 4 pixel pipelines with 2 TMUs each.
In fact, Xabre 600 once again surprises us with its rather high results: it managed to settle right after GeForce FX 5800 Ultra, Radeon 9700 Pro, GeForce4 Ti4600 and Ti4800, having left all other competitors behind. Unfortunately this card didn’t prove so successful in gaming tests. Let’s figure out why it happened like that. Just like GeForce4 Titanium GPUs, SiS Xabre600 has four pixel pipelines with two texturing units each. The graphics core works at 300MHz, so Xabre 600 is similar to GeForce4 Ti4600 in texture rendering. It only has lower memory frequency: 600MHz against 650MHz by GeForce4 Ti4600/4800. So, it’s quite natural that Xabre 600 performs well in multi-texturing. But why don’t we see these impressive results in single-texturing mode of Fillrate tests? The reason for Xabre 600 doing worse here is its slower memory controller. Single-texturing requires more active communication with the graphics memory than in case of multi-texturing (Each time the graphics chip reads pixel color from the graphics memory, then lays the color of the nearest transparent pixel over it. The operation is repeated 63 times for all 64 flat surfaces.). In other words, higher memory workload pushes Xabre 600 to lose to GeForce4 Titanium series that has faster doubled memory controllers.
Another interesting pair of chips is RADEON 9100 and RADEON 9000. In Single-texturing mode RADEON 9000 with 250MHz chip frequency outpaces RADEON 9100 with 275MHz chip frequency. The memory of RADEON 9100 works at 550MHz, while that of RADEON 9000 – at 500MHz. Therefore, the only reason for it to be faster than RADEON 9100 in Single-texturing mode may be only connected with the optimization of RADEON 9000 as this graphics chip was released after 9100. And in case of multi-texturing RADEON 9100 performs faster as it has two texturing units per pixel pipeline, while RADEON 9000 has only one.
Summing it up, we should say that the results displayed in the diagram correspond to what the manufacturers claim in the official specs.
This test requires hardware vertex shaders support. GeForce4 MX440-8x executes them via software and thus is excluded from this test. Xabre 600 also executes vertex shaders at the software level (Vertexlizer technology), but it managed to cheat 3DMark03 and run the test :) .
So, what do we see? This is one more test where RADEON 9700 PRO is the leader. Well, judging by the tests, RADEON 9700 PRO, 9500 PRO and 9500 work very well with vertex shaders. Other graphics cards can’t compete with the two latest ATI’s solutions, except GeForceFX 5800 Ultra. Well, I can congratulate ATI fans with this victory :)
The sixth position after the four RADEON 9x00 based graphics cards and GeForceFX is assigned to GeForce4 Ti4600 and GeForce4 Ti4800. Both graphics cards run equally fast, showing a bit higher result than RADEON 9100. and this is again very easy to explain if we calculate the number of pipelines in every chip responsible for vertex shaders implementation. GeForce4 Titanium solutions feature two units like that, while RADEON 9700 PRO, 9500 PRO and 9500 – four. That is why GeForce4 GPUs lose to the more advanced RADEONs here.
Xabre 600 with its software vertex shaders execution can’t even reach GeForce4 Ti4200-8x, although we used a powerful CPU – Pentium 4 2.8GHz.
GeForce3 Ti200 proves the slowest in this test. It’s quite natural as this was the first GPU from NVIDIA to feature hardware vertex shaders support. It has only one vertex pipeline and works at the lowest frequencies of all participating solutions: 175MHz.
Overall, “Vertex Shader” test seems to be a very accurate tool for measuring vertex shaders execution speed. All the results we’ve got comply with the theoretical presuppositions pretty much.
This test measures the system performance during the execution of ver.2.0 pixel shaders. Of course, this test can only run on cards supporting DirectX9. So, we will take a look at the results shown by GeForce FX 5800 Ultra, Radeon 9700 PRO, RADEON 9500 PRO and RADEON 9500.
Well, NVIDIA has still to improve the GPU and drivers. GeForceFX 5800 Ultra loses here to all R300 based solutions from ATI: even to 64MB RADEON 9500. A triumph of ATI, no doubt :).
This test is new to 3DMark03. It shows how well the system can distribute the workload between the CPU and graphics card Let’s check the results now.
And once again, ATI wins the test. RADEON 9700 PRO and 9500 PRO outperform GeForceFX 5800 Ultra, while 64MB RADEON 9500 kills GeForce4 Titanium GPUs. Xabre 600 is again in the end. This is quite an expected picture considering the complexity of the scene and the fact that this graphics solution emulates vertex shaders on the software level.
So, we have tested thirteen graphics cards ranging from the value GeForce3 Ti200 and RADEON 9100 to the latest newcomer GeForceFX 5800 Ultra. Of course, we can’t make any far-reaching statements basing on results of one benchmarking package, but anyway there are some things to be mentioned.
First of all, we should congratulate NVIDIA and the company’s software developers. Their 42.68 drivers helped GeForceFX 5800 Ultra based card to outrun its main rival, RADEON 9700 PRO nearly in all the tests and resolutions. We won’t say anything about the price-to-performance ratio of the two leaders in the current graphics card market, as the price of GeForceFX 5800 Ultra has not yet been finalized. You can only be sure that it’s going to be VERY expensive.
We would also like to single out RADEON 9500 PRO based card. It showed really worthy performance in all gaming and theoretical tests following the two leaders: RADEON 9700 PRO and GeForceFX 5800 Ultra. The two RADEON 9500 based cards showed that the amount of graphics memory as well as the width of the graphics memory bus doesn’t always affect the performance much. These two behave much alike in some tests.
Graphics cards based on GeForce4 Titanium chips are not to be dismissed as obsolete yet. They are quite powerful gaming chips, unfortunately without ver.1.4 pixel shaders and ver.2.0 vertex shaders support. That’s why graphics cards based on GeForce4 Ti4200-8x, Ti4600 and Ti4800 GPUs perform well in the first gaming test, “Wings of Fury”, which doesn’t use pixel shaders, and in some theoretical tests. But when it comes to complex Game 2 and 3 tests, they lose to RADEON 9500 / 9500 PRO / 9700 PRO based cards. There is nothing to be done about it, unfortunately.
Xabre 600 is quite a different story. The main advantage of this graphics card is that it rose to the occasion and managed to run 3DMark03 tests. We can’t take seriously the results of the CPU tests where Xabre 600 outruns both: RADEON 9700 and GeForceFX 5800 Ultra. This is no more than an error (or peculiarity?) of the benchmarking software. With all its artifacts, like displaying just one half of the screen, Xabre 600 shouldn’t be allowed to run 3DMark03 at all.
As for the remaining cards, like GeForce4 MX440-8x, GeForce3 Ti200, RADEON 9100 and 9000, there is only one test in 3DMark03 for them: “Game 1 – Wings of Fury”. In complex tests, like Game 2 and 3, they show too low performance. It’s not correct to compare them to other, more advanced graphics cards. Futuremark Company claims even that Game 1 is the only test intended for cards like that.
Summing all up, we should say that the 3DMark benchmarking set is too “new” and too “heavy” for most graphics cards now available. Still, as our experiments showed, it can be used to compare performances of modern DirectX9 graphics chips. We guess this benchmarking tool is going to become a standard de facto in graphics cards benchmarking, like 3DMark2001 SE has been hitherto.