by Tim Tscheblockov
03/16/2004 | 05:22 PM
Each of you has to face the choice of the proper graphics card for the home system sooner or later. Today the leaders of the 3D market are R3xx from ATI and NV3x from NVIDIA. The graphics cards based on top chips, such as ATI RADEON 9800 XT and NVIDIA GeForce FX 5950 Ultra cost quite a bit of money but guarantee that all contemporary games will run without any problems on them. The cards from a better value price group, such as RADEON 9600/9600 PRO/9600 XT and NVIDIA GeForce FX 5600/5600 Ultra/5700/5700 Ultra cost a little less but at the same time offer lower performance.
But our life would be boring if the graphics card choice were limited by these few alternatives only. In order to introduce a bit of an intrigue, the manufacturers enriched their product range with the whole bunch of low-cost modifications: RADEON 9800 SE, GeForce FX 5900LE/XT and the like, so that it turns out really hard now to figure out if the card costs the money they ask for it or not.
In our today’s article we will take a closer look at the top-end graphics cards and their low-cost modifications, will pay due attention to their overclocking potential and their ability to be turned into a “fully-fledged” solution, and finally test them in all sorts of contemporary games in order to find out the best choice for your needs.
The today’s most powerful graphics processor in the ATI product family will be represented by ASUS A9800XT/TVD.
Top-end ASUS graphics cards usually follow the reference design and differ from the standard solutions by their higher functionality, such as the implemented hardware monitoring, for instance. RADEON 9800 XT based graphics card we are talking about now is also no exception.
This graphics monster from ASUS looks very neat and even miniature unlike the actual reference solution from ATI. The cooling system for the graphics processor and the memory chips consists of two parts. The cooler on the front of the PCB is designed on a copper footing with a heat pipe, aluminum heatsinks screwed to it and two fans. On the reverse side of the PCB there is a copper plate taking the heat off the memory chips. The important thing about this cooling system from ASUS is that it takes not too much room and allows installing any other cards into the PCI slot located next to the AGP.
The graphics card is equipped with 256MB of DDR SDRAM memory from Hynix with 2.5ns cycle time (Hynix HY5DU573222 AFM-25). The graphics processor and the memory of ASUS A9800XT/TVD work at 412MHz and 730MHz respectively.
We didn’t overclock this graphics card during our test session. Why? This is very simple. First, all juices have already been squeezed from RADEON 9800 XT, so that only extreme overclocking can ensure any tangible performance growth. And second, graphics cards of the kind are very rarely overclocked, because they are the top solutions in the product range and their owners have nothing to strive for :)
ATI RADEON 9800 PRO (R350) used to be the fastest graphics chip from ATI until ATI RADEON 9800 XT (R360) arrived. In fact, RADEON 9800 XT is none other but an officially overclocked RADEON 9800 PRO.
When RADEON 9800 XT took over the title of the ATI performance leader from the predecessor, the prices of the solutions based on “outdated” ATI chips went down quite noticeably. The today’s pricing situation allows positioning NVIDIA’s GeForce FX 5900 as a competitor to RADEON 9800 PRO and not RADEON 9800 XT any more.
This graphics chip in our today’s roundup will be represented by a graphics card from Sapphire:
Well, there is nothing extraordinary I should say: the card follows RADEON 9800 PRO reference design, features a standard cooling system onboard, 128MB of DDR SDRAM graphics memory from Samsung with 2.8ns cycle time (SAMSUNG K4D263238e-GC2A). The graphics processor and the graphics memory work at the nominal frequencies of 380MHz and 680MHz respectively.
I wouldn’t call the overclocking potential of this card very impressive: the graphics card managed to work at the maximum of 420MHz for the chip and 720MHz for the memory. As a result, the overclocked RADEON 9800 PRO graphics core appeared only a little bit faster than the RADEON 9800 XT graphics core working at its nominal frequency, while the graphics memory of the PRO version was only 10MHz behind the graphics memory of the XT.
The story of ATI RADEON 9800 SE based graphics cards is very similar to the story of RADEON 9700 and their cut-down modifications, RADEON 9500. Although this time, it is not R300, but R350, which loses some of its functional units. So the resulting graphics chip has not a different numeric index like in case of RADEON 9700 and 9500, but the same 9800 index though with the “SE” suffix in the end.
There are only three differences between the RADEON 9800 SE and the fully-fledged RADEON 9800/9800 PRO: two grave differences and one trifling difference:
In fact, RADEON 9800 SE is based on the chips, which failed the test for “fully-fledged” chips, but are perfectly fit for work with 4 disabled pixel pipelines and disabled HyperZ III. I would like to point out that physically functional chip units are not disabled in this case. Instead they change the DeviceID of the graphics processor, and when the graphics card driver sees the RADEON 9800 SE DeviceID, it doesn’t activate half of the available pixel pipelines and HyperZ. When we change the DeviceID of the chip on the hardware or software level, so that it again indicates RADEON 9800/9800 PRO, the driver starts using the graphics processor potential to the full extent.
Of course, when we transform RADEON 9800 SE into RADEON 9800 far not all the graphics chips remain operational. In case of a failure, you may notice some visual artifacts caused by incorrect HyperZ III functioning. Sometimes, they may also result from the deficiencies of the previously disabled pixel pipelines. “Successful” modification implies that the bugs discovered during the quality tests on the fab are either insignificant, or have no effect on the image quality, or pop up in some rare exotic conditions.
RADEON 9800 SE based graphics cards are positioned as low-cost versions of the top RADEON 9800 and follow the reference design of RADEON 9700 and 9500 PRO instead of that of the RADEON 9800/9800 PRO based solutions. Therefore, there are two major types of graphics cards known as RADEON 9800 SE: those based on RADEON 9700 reference design and equipped with 256bit memory bus and those on RADEON 9500 PRO reference design equipped with 128bit memory bus.
The graphics card modification with 256bit bus will be represented by a Sapphire solution:
Actually the graphics processor is the only thing that distinguishes this graphics card from RADEON 9700. all the other elements of the card remained absolutely the same: even the cooler is the same as those used on RADEON 9700/9700 PRO based graphics accelerators.
The card is equipped with 128MB of DDR SDRAM from Infineon with 3ns cycle time (Infineon HYB25D128323C-3.3). The graphics processor works at 325MHz and the graphics memory – at 550MHz frequency.
To perform the necessary modifications I used a special PatchScript utility included into the Rivatuner package and after that the card worked flawlessly as a RADEON 9800. However, its overclocking potential was far from impressive: the graphics chip and memory managed to retain stability at only 360MHz and 600MHz respectively.
These pretty low overclocking results can partially be explained by the fact that the power voltage of the graphics processor on this card equals the nominal voltage of the RADEON 9700, since the card is designed according to the RADEON 9700 reference layout. And this value is far from the acceptable voltage for RADEON 9800, because it equals only 1.54V instead of the required 1.7V.
Graphics cards based on RADEON 9800 SE and equipped with 128bit memory bus are based on RADEON 9500 PRO reference design. Today we will take a look at the card from Sapphire representing this item of ATI product range.
This graphics card differs from RADEON 9500 PRO only by the stickers and the cooler shape: coolers like that are usually installed onto All-In-Wonder RADEON 9700 PRO based graphics cards. To tell the truth, I got the impression that they assembled this least expensive graphics card from the RADEON 9800/9800 SE family of the parts that were left in the warehouse. :)
The card is equipped with 128MB graphics memory from Samsung with 3.3ns cycle time (Samsung K4D263238E-GC33). The nominal working frequencies of the graphics processor and graphics memory appeared lower than those of RADEON 9800 SE 256bit. Namely, they were 275MHz for the chip and 540MHz for the memory.
RADEON 9800 SE with 128bit memory bus can also be modified, i.e. it is possible to enable all 8 pixel pipelines and HyperZ III. However, if the modified graphics cards with 256bit memory could be called fully-fledged RADEON 9800 solutions, then the 8-pipeline RADEON 9800 with 128bit memory bus do not exist in the market and there is no DeviceID for them.
Anyway, this fact will never stop us from enabling everything possible in the RADEON 9800 SE 128bit with the help of Rivatuner utility and obtaining a RADEON 9800 with 128bit memory access. In fact, this is almost a complete analogue of the RADEON 9500 PRO, which also features 8 pixel pipelines and 128bit memory bus.
I have to confess that the modification of the graphics card we had at our disposal appeared not absolutely successful. When we turned our chip into RADEON 9800, i.e. when we enabled all 8 pixel pipelines and HyperZ III, we noticed a missing image square at the top of the screen approximately 8x8 pixels big in all games. This is direct evidence that this graphics processor was assigned for RADEON 9800 SE 128bit solution because of the defective part of cache for the low-resolution Z-buffer. Besides this missing square we revealed no other unpleasant surprises throughout the entire test session.
The incorrect display of 64 pixels hardly told on the image processing speed (I even took the trouble of calculating that 64 pixels occupy only 0.008% of the screen area for 1024x768 resolution). That is why I decided that there is no need to replace this graphics card with a different one and disregarded the revealed artifact this time, especially since we are going to consider the performance and not the image quality that much today.
The overclocking of this graphics card appeared really pleasing: the maximum chip and memory frequencies when the card remained stable equaled 390MHz for the chip and 700MHz for the memory, which means that we managed to speed up the core by 42% and the graphics memory by a little less than 30%. This is an excellent result, I should say!
But let’s not make any preliminary announcements: I suggest that we find out first how greatly the performance of this highly overclockable “new RADEON 9500 PRO” solution will be limited by the narrow 128bit memory bus.
The last one in the list of ATI graphics cards tested today comes RADEON 9600 XT, which we will consider the fully-fledged representative of the mainstream price category and a direct alternative to RADEON 9800 SE.
Again we will take a graphics card from Sapphire:
Unlike RADEON 9700/9800 chips, this graphics processor was initially developed as a solution for mainstream graphics accelerators that is why it features less complex architecture than the RADEON 9700/9800. RADEON 9600 chips are provided with only 4 pixel pipelines and two vertex processors. I would also like to point out one more distinguishing feature of the RADEON 9600 family: they are manufactured with finer 0.13micron production technology. As a result, RADEON 9600/9600 PRO/9600 XT consume less power and generate less heat than RADEON 9700/9800 solutions.
All this is proven by the looks of the Sapphire card based on RADEON 9600 XT: it has no connector for additional power supply cable, and the graphics processor is cooled down by a pretty weak device.
The fine manufacturing technology also influences the performance of the graphics chip: its nominal working frequency has been set at a level, which is usually unattainable for the ATI chips manufactured with 0.15micron process (without extreme overclocking tricks). The nominal working frequency of the RADEON 9600 XT chip equals 500MHz.
The graphics processor features 128bit memory access. The sapphire card is equipped with DDR SDRAM chips from Samsung in TSOP packages with 3.3ns cycle time (Samsung K4D551638D-TC33). This memory works at 600MHz nominal frequency.
During overclocking the working frequencies of this graphics card grew up quite significantly, and reached 625MHz for the chip and 740MHz for the memory. This is a very nice result, I should say, for both: the graphics processor, which already boasted the highest clock frequency of all ATI chips, and for TSOP-memory chips, which maximum working frequency is exactly 600MHz as the marking suggests.
NVIDIA’s graphics chips leader, NVIDIA GeForce FX 5950 Ultra, will be represented by a VGA card from Leadtek:
Leadtek WinFast A380 Ultra follows the reference design for NVIDIA GeForce FX 5950 Ultra, but unlike the reference cards with their traditional cooling solution, which sucks the air in from outside the case, Leadtek uses its own cooler of specific construction and hence doesn’t require a double bracket for the graphics card fastening. Nevertheless, you will not be able to install any PCI cards in the next PCI slot, because Leadtek’s cooler is quite tall and even if you fit the PCI card in, it will block the air flow completely for the graphics card.
Leadtek WinFast A380 Ultra cooling solution consists of two parts: there is a copper cooler on the front side of the PCB, which covers the GPU and the memory chips at the same time, and an aluminum heatsink on the reverse side of the PCB, which dissipates the heat coming from the memory chips installed there.
The card is equipped with 256MB of graphics DDR SDRAM from Hynix with 2ns cycle time (Hynix HY5DU283222 AF-2). The clock frequency of the graphics processor and graphics memory equals 475MHz and 950MHz respectively.
Graphics cards based on NVIDIA GeForce FX 5900 are none other but low-cost versions of the GeForce FX 5900 Ultra based solutions, which were the top ones before GeForce FX 5950 Ultra arrived.
GeForce FX 5900 Ultra and GeForce FX 5900 are based on the same PCB layout, which is very similar to that of GeForce FX 5950 Ultra. The differences can be noticed only in the voltage regulator circuitry.
The graphics cards based on GeForce FX 5900 have two major differences from the Ultra modifications. First, they feature only 128MB of memory instead of 256MB: the memory chips are only on one side of the PCB in this case. Second, the GPU working frequency in 3D mode is reduced to 400MHz from 450MHz.
NVIDIA GeForce FX 5900 will be represented by ASUS V9950 graphics card:
ASUS graphics card follows the reference design for NVIDIA GeForce FX 5900/5900 Ultra, but it features only 128MB of memory as it is a non-Ultra version. The card uses DDR SDRAM from Hynix with 2.2ns cycle time (Hynix HY5DU283222 AF-22). The memory chips are installed on the front side of the PCB and are covered with a cooler together with the CPU. This cooler consists of a copper footing with the ribs soldered to it and two fans blowing the air through the heatsink ribs.
ASUS graphics card carefully follows NVIDIA’s recommendations regarding the working frequencies of the chip and memory: 400MHz and 850MHz in 3D mode respectively.
When we overclocked the card it managed to work stably at 480MHz core and 980MHz memory frequencies, which is a pretty good result for both: the graphics processor and the memory. The overclocked card appeared even faster than NVIDIA GeForce FX 5950 Ultra at nominal frequencies (475MHz/950MHz). Later we will see if this frequency gain is enough for the solution to outperform the today’s gaming leader from NVIDIA.
“XT” suffix by the graphics cards based on NVIDIA processors means just he opposite thing compared to what the same “XT” by ATI stands for. It is probably in order to “please” ATI that NVIDIA marked its slower GPU modifications with the “XT” suffix, while ATI uses it to mark the faster VPU modifications, as we all know.
GeForce FX 5900 XT is a lower-cost modification of GeForce FX 5900.
Unfortunately, we failed to find out what company produced this graphics card, however, we suspect that it was the Chinese Colorful Company. The PCB layout of this graphics solution looks very similar to that of other GeForce FX 5900 XT based products.
In fact, the components layout is very much different from what NVIDIA suggests with its GeForce FX 5900/5900 Ultra, as our hero is a low-cost modification. For instance, all memory chips are placed on one end of the PCB, which made the wiring much simpler. The chips are slower and hence cost less than those memory chips they use for GeForce FX 5900/5900 Ultra. The memory chips are located not in a half-circle, but in two rows, which allows using simple square heatsinks for their cooling. Finally, the voltage regulators for the GPU and graphics memory do not look that monstrous any more as those we saw on NVIDIA GeForce FX 5950 Ultra or GeForce FX 5900/5900 Ultra. Probably lower working frequencies of the GPU and graphics memory allow reducing the requirements to their voltage regulators.
As a result, we get a very nice product. Firstly, this card is pretty quiet and its cooling system doesn’t occupy the space around the next PCI slot. And secondly, the card is equipped with a fully-fledged NV35 graphics processor and 256bit memory bus, and its working frequencies are not that much lower than those of the top-end products.
The GPU and the graphics memory of our graphics card work at 400MHz and 700MHz respectively. The card is equipped with 128MB of DDR SDRAM graphics memory in8 BGA chips from Hynix with 2.8ns cycle time (Hynix HY5DU283222 AF-28).
After overclocking the GPU and the memory of our graphics card worked at 440MHz and 860MHz respectively.
GPU appeared very poorly overclockable, which was quite a surprise, as I expected it to reach at least 450MHz, as GeForce FX 5900 XT, 5900 and 5900 Ultra are physically the same. GeForce FX 5900 XT is very unlikely to be a defective GeForce FX 5900, i.e. NV35 chip, which is not capable of working at a certain frequency, because the nominal working frequency of the XT solution is the same as that of the GeForce FX 5900. However, when we overclocked the card to 440MHz it refused to work stably. Maybe I just was unlucky and got not the best overclockable piece, or maybe I had to use a better cooling system to achieve higher results.
The graphics memory, on the contrary, appeared highly overclockable: the maximum working frequency of these memory chips is 714MHz, as the marking suggests, so the 860MHz we managed to squeeze out of this memory is an excellent result, I should say.
NVIDIA GeForce FX 5700 Ultra is the fastest mainstream representative in our today’s test session. It gets really close to the slowest NV35 based solutions from both: price and performance point of view. This GPU will be represented by an NVIDIA reference card:
The card is equipped with 128MB of DDR-II graphics memory from Samsung with 2ns cycle time (Samsung K4N26323ae-GC1K). The GPU works at the nominal frequency of 475MHz and the memory – 900MHz.
The card doesn’t boast any superfluous overclockability: the maximum we managed to make it work at was 525MHz for the chip and 1000MHz for the memory.
Now that we have introduced to you all our testing participants, which will compete for the title of the “Best Overclocker’s Choice”, let’s pass over to the actual benchmarks.
Our test system was configured as follows:
We used the following software:
We tested all graphics cards with two types of image quality settings. In both cases we set the graphics quality in the driver to the maximum (these are the settings responsible for the texture quality in Catalyst driver and the “speed-quality” slider in ForceWare panel). In the first type of settings we didn’t enable full-screen anti-aliasing and anisotropic filtering, while in the second we forced 8x anisotropic filtering (in Quality mode) and 4x FSAA.
In order to eliminate the influence of the side factors on the VGA cards performance, we tested all of them with disabled sound, except DeusEx 2 game, which refused to run without sound. Vsync was disabled in all cases.
So, let’s get started!
Judging by the demo versions I could say that Far Cry is the freshest, cutest and probably the most demanding to graphics of all 3D shooters, which have already been released. To create the beautiful landscapes of the tropical island the gaming engine of Far Cry Demo uses DirectX shaders very actively.
The engine involves pixel and vertex shaders version 1.1 complying with DirectX8 specification and pixel and vertex shaders DirectX9 2.0 for graphics cards on ATI chips and 2.x for graphics cards on NVIDIA chips.
The consequences of this difference in the shaders used for various graphics cards is not so evident: my personal experience suggests that you can actually see the visual difference between shader version 2.0 and 2.x only when we have small ripples on the water surface and see the lake floor through them.
Forced FSAA and anisotropic filtering in the driver control panel leads to incorrect image display in Far Cry Demo that is why we didn’t run the tests with maximum image quality. We measured the graphics cards performance in the regular mode with the graphics quality settings of Far Cry Demo at the maximum:
You can clearly see that the graphics processor performance in Far Cry Demo affects the results much more than the memory bus bandwidth. In fact, this is not at all surprising, taking into account how actively Far Cry Demo uses shaders. For instance, RADEON 9800 SE with narrow 128bit memory bus didn’t fall far behind RADEON 9800 SE with 256bit memory bus. Moreover, this situation remained the same even after the modifications.
All in all, I can conclude that ATI based graphics cards are overall better here than the solutions based on NVIDIA chips: ATI based solutions use shaders more actively.
As a result, ATI RADEON 9800 XT and 9800 PRO appeared ahead of NVIDIA GeForce FX 5950 Ultra and GeForce FX 5900. the modified RADEON 9800 SE outperformed the overclocked NVIDIA GeForce FX 5900 XT, and RADEON 9600 XT turned out faster than GeForce FX 5700 Ultra.
The modification and overclocking of the RADEON 9800 SE based graphics cards almost doubles their performance in far Cry Demo and makes them catch up with NVIDIA GeForce FX 5950 Ultra/5900/5900 XT. Even the card with a 128bit memory bus, which sometimes gets missed out, performed brilliantly after overclocking and corresponding modding.
However, we shouldn’t forget that far not all RADEON 9800 SE based solutions will successfully turn into a fully-fledged RADEON 9800. Without the enabled 8 pixel pipelines and HyperZ III RADEON 9800 SE based graphics cards have nothing to beat the mainstream RADEON 9600 XT with: even at nominal frequencies it outperforms the top-end “cuts”.
The situation with NVIDIA based graphics cards is just the opposite: the “cut-down” GeForce FX 5900 XT easily outpaced GeForce FX 5700 Ultra even without any modifications, and when overclocked it manages to leave behind even GeForce FX 5900.
Halo uses DirectX8 vertex shaders version 1.1 and pixel shaders version 1.1, 1.4 and 2.0. The game doesn’t support full-screen anti-aliasing, because it builds the image not in the frame buffer but as a texture in order to be able to apply all special effects. Therefore we ran all the benchmarks with the gaming settings at the maximum graphics quality, but without forced anisotropic filtering and full-screen anti-aliasing.
The results of the tested graphics cards in 1024x768 and 1280x1024 are limited by the processor speed and the overall system performance, however, in 1600x1200 this influence disappears. Nevertheless, in 1600x1200 most graphics cards perform equally fast. Only RADEON 9800 SE working at nominal frequencies, RADEON 9600 XT and NVIDIA GeForce FX 5700 Ultra fall behind the leading group.
The top-end graphics cards on ATI and NVIDIA chips show almost the same performance, so it is really hard to single out a leader. The modified RADEON 9800 SE cards yield to the performance leaders, while GeForce FX 5900 XT manages to stay in the leading group even when it doesn’t undergo any overclocking.
Due to higher nominal clock frequency of the VPU as well as working HyperZ III technology, ATI RADEON 9600 XT manages to leave RADEON 9800 SE far behind and runs almost as fast as NVIDIA GeForce FX 5700 Ultra or even a little faster.
Tomb Raider: Angel of Darkness uses pixel shaders version 1.1, 1.4 and 2.0 and vertex shaders 1.1 and 2.0. During the tests I sued one of the “heaviest” scenes – paris5_4. The graphics quality settings in the game were adjusted according to the defaults for PS 2.0 mode, Vsync and Pixel Shader 2.0 Shadows were disabled. Because of the incorrect work of FSAA, we didn’t force anisotropic filtering and full-screen anti-aliasing.
The results obtained in this gaming application are almost fully determined by the pixel shader efficiency and the graphics processor clock frequency. However, this is not always like that. For example, the modified ATI RADEON 9800 SE with 128bit memory bus failed to outperform the modified RADEON 9800 SE with 256bit memory bus, even despite the higher working frequencies of the VPU and memory of the former.
All in all, ATI based graphics cards are ahead in paris5_4 scene due to their more efficient pixel shaders. In this scene pixel shaders are very actively used, for example for such effects as washed out objects out of the camera focus or light refraction when it goes through the stream of heated air.
The overclocked ATI RADEON 9800 PRO outperforms RADEON 9800 XT. Overclocked NVIDIA GeForce FX 5900 outpaces GeForce FX 5950 Ultra, and the modified RADEON 9800 SE and GeForce FX 5900 XT are just a little bit behind the leaders.
The mainstream RADEON 9600 XT runs almost neck and neck with GeForce FX 5900 XT and proves faster than RADEON 9800 SE and GeForce FX 5700 Ultra working at their nominal frequencies.
Deus Ex 2 is based on a modified Unreal Tournament engine using pixel and vertex DierctX8 shaders. The modifications introduce such spirits of the times as dynamically calculated shadows and full-screen effects of image post-processing. When these effects are enabled (Bloom filter) the game doesn’t support FSAA, but the filter smoothens the entire scene in general and the polygon borders in particular.
The graphics cards tests in Deus Ex 2 were carried out with maximum graphics quality settings, i.e. with enabled Bloom but without forced full-screen anti-aliasing and anisotropic filtering.
Just like in Tomb Raider: Angel of Darkness the graphics cards on ATI chips took the lead.
Overclocked RADEON 9800 PRO made it the winner in this benchmark and the modified and overclocked RADEON 9800 SE demonstrated more than twofold performance growth. When these cards are working as it, i.e. at their nominal frequencies, they yield significantly to RADEON 9600 XT.
The overclocked NVIDIA GeForce FX 5900 is a little bit faster than GeForce FX 5950 Ultra, and the overclocked GeForce FX 5900 XT outperforms GeForce FX 5900 and appears only a little bit behind GeForce FX 5950 Ultra.
GeForce FX 5700 Ultra fails to compete with RADEON 9600 XT even after successful overclocking, not to mention the slow GeForce FX 5900 XT.
Unreal Tournament 2004 graphics engine didn’t undergo any changes compared with Unreal Tournament 2003 that is why we will not see any pixel shaders in this game. During the tests we used maximum graphics quality settings available in the demo version of the game.
The results are mostly determined by the fillrate and the memory bus bandwidth.
As we see, the overclocked ATI RADEON 9800 PRO is a little ahead of RADEON 9800 XT.
RADEON 9800 SE with 128bit memory bus yields to its 256bit analogue in both cases: with and without overclocking-friendly modifications, despite the higher working frequencies of the graphics processor and memory its performance is strongly affected by the narrow 128bit memory bus.
Due to successful overclocking and high working frequencies, RADEON 9600 XT defeats the 8-pipeline RADEON 9800 SE with 128bit memory bus, however, the 256bit version of the card appeared over its head.
Among NVIDIA based graphics cards, the leadership belongs to the overclocked GeForce FX 5900. GeForce FX 5900 XT is just a little bit slower than GeForce FX 5900 and when overclocked, it gets really close to GeForce FX 5950 Ultra.
GeForce FX 5700 Ultra based graphics card working at nominal frequencies outperforms ATI RADEON 9600 XT, but its poor overclocking potential makes it yield to ATI RADEON 9600 XT. And as for GeForce FX 5900 XT, there is nothing to comment on.
With enabled full-screen anti-aliasing and anisotropic filtering the overall picture remains the same.
But look, how greatly grew up the results shown by the modified RADEON 9800 SE! The improvement is almost triple!
This performance gain results from a combination of three factors. Firstly, enabling 8 pixel pipelines doubles the fillrate. Secondly, overclocking improves the fillrate and the memory bus bandwidth in proportion to the graphics processor and memory speed increase. Finally, enabled HyperZ III increases the efficiency of the available memory bus bandwidth, which is especially valuable for modes with enabled FSAA.
Here I would like to draw your attention to the remarkable behavior of RADEON 9600 XT based solution: if in 1024x768 and 1280x1024 this card manages to run faster than RADEON 9800 SE, which gets a fully-fledged RADEON 9800 chips after the modifications, then in 1600x1200 RADEON 9600 XT yields notably to the rival. It looks as if RADEON 9600 XT featured a smaller low resolution Z-buffer cache, just like all other RV350/RV360 based products. So, in 1600x1200 with enabled FSAA 4x this buffer doesn’t fit into the cache, which automatically disables HyperZ III by RADEON 9600 XT and hence the performance drops.
Max Payne 2 uses DirectX8.1 pixel shaders version 1.1 and 1.45 (DirectX9 compatible cards of course, support pixel shaders 1.4) and vertex shaders version 1.1. We ran the tests in this game in two modes: with maximum gaming graphics quality settings but without anisotropic filtering and full-screen anti-aliasing, and with these functions forced in the drivers.
The results demonstrated by our testing participants in Max Payne appeared limited by the processor and overall system performance. Only in 1600x1200 this limiting influence gets so insignificant that we get the opportunity to estimate the graphics cards performance in this game.
ATI RADEON 9800 XT and RADEON 9800 PRO showed almost identical results: their performance appeared limited by the CPU and the overall system performance.
The modified RADEON 9800 SE based graphics cards managed to catch up with NVIDIA GeForce FX 5950 Ultra and overclocked GeForce FX 5900/5900 XT, while the same cards working at the nominal frequencies didn’t beat even RADEON 9600 XT and GeForce FX 5700 Ultra.
NVIDIA GeForce FX 5900 and GeForce FX 5900 XT showed almost the same results with and without overclocking, which is another point in favor of a low-cost GeForce FX 5900 XT as an “Overclocker’s choice”.
With enabled full-screen anti-aliasing and anisotropic filtering 128MB graphics cards on NVIDIA chips lacked graphics memory to be able to work in 1600x1200 resolution that is why you see the results only for 1024x768 and 1280x1024 modes.
The results show very clearly that forced FSAA requires much higher data transfer rate along the graphics memory bus and RADEON 9800 SE with its 128bit memory access falls considerably behind its counterpart with 256bit bus even though the 128bit version boasts higher working frequencies than the 256bit one.
You should also take a closer look at the results obtained by NVIDIA GeForce FX 5950 Ultra in 1280x1024 resolution: this graphics card appeared far ahead all other testing participants on NVIDIA chips here. It most probably owes this victory to the 256MB of onboard graphics memory, while all other NVIDIA based graphics cards had to transfer the textures via a relatively slow AGP bus.
Prince of Persia: Sands of Time is a 3D reincarnation of the ancient Prince of Persia, which turned out a worthy continuation of the legend in terms of style, graphics and the gaming process. The gaming engine uses DirectX8 pixel and vertex shaders, which you can see from the picture, actually: water surfaces in the game refract the light the right way, the camera receives the fully-fledged Motion Blur when moving fast enough, and the brightly lit objects get a typical halo effect (similar to Bloom from Deus Ex 2 and full-screen effects from Tron 2.0 or Splinter Cell). It is especially pleasant that all these nice things get along very well with full-screen anti-aliasing that is why we tested our graphics cards in two modes: without the full-screen anti-aliasing and anisotropic filtering and with these functions forced in the drivers.
Here NVIDIA based graphics cards perform much better than the solutions based on ATI graphics processors: the game doesn’t use DirectX9 pixel shaders.
Due to higher VPU frequency and just a little lower graphics memory frequency the overclocked ATI RADEON 9800 PRO manages to slightly outperform RADEON 9800 XT.
The modified RADEON 9800 SE gets close to RADEON 9800 PRO, but fails to outperform it because of the lower memory bus bandwidth.
Higher working frequencies of RADEON 9600 XT let it outperform RADEON 9800 SE based cards running at their nominal frequencies. And when overclocked, RADEON 9600 XT even gets close to the 8-pipeline RADEON 9800 SE with 128bit memory bus.
NVIDIA GeForce FX 5900 and 5900 XT show almost the same results. After successful overclocking, GeForce FX 5900 outperforms GeForce FX 5950 Ultra, and GeForce FX 5900 XT yields a little bit to the latter.
NVIDIA GeForce FX 5700 Ultra defeats RADEON 9600 XT, but it is certainly unable to catch up with a GeForce FX 5900 XT based graphics card equipped with a fully-fledged NV35 chip.
With enabled full-screen anti-aliasing and anisotropic filtering the overall picture remains unchanged.
NVIDIA GeForce FX 5900 XT and ATI RADEON 9800 PRO working at their nominal clock rates yield a little bit to the top-end graphics solutions, while after successful overclocking they perform almost as fast as RADEON 9800 XT and GeForce FX 5900 Ultra.
RADEON 9800 SE based graphics cards benefit even more from modifications and overclocking: now the cards get fantastically fast showing almost three times the performance in 1600x1200 resolution.
RADEON 9600 XT again shows a noticeable performance drop in 1600x1200. In lower resolutions it yielded just slightly to NVIDIA GeForce FX 5700 Ultra, but in 1600x1200 the gap between them got almost double.
The game Star Wars: Knights of the Old Republic came from XBOX RPG platform and boasts a graphics engine of an excellent Action game. The gaming engine works in OpenGL and used special extensions for vertex and pixel tools similar to DirectX8 pixel and vertex shaders. The graphics cards were tested in two modes in this game: without the full-screen anti-aliasing and anisotropic filtering and with these functions forced in the drivers.
The fillrate and the graphics memory bus bandwidth are the No. 1 determinatives for the graphics cards results in this gaming application.
For instance, the overclocked RADEON 9800 PRO boasts a slightly higher core clock frequency than RADEON 9800 XT and in 1024x768 and 1280x1024 resolutions it outperforms the rival. However, as the resolution grows, the dependence on the memory performance increases and in 1600x1200 the results of RADEON 9800 XT and RADEON 9800 PRO with slightly lower frequency of the overclocked graphics memory level out.
The same way, RADEON 9800 SE with 128bit memory bus yields to RADEON 9800 SE with 256bit memory bus despite the higher working frequencies of the former. And the higher grows the resolution, the bigger turns the gap between the competitors.
RADEON 9600 XT outperforms the nominal RADEON 9800 SE even without any overclocking. However, when we modify the SE cards, 9600 is no longer able to compete with them even after overclocking.
The non-overclocked NVIDIA GeForce FX 5900 and 5900 XT show almost similar results and when overclocked, GeForce FX 5900 catches up with GeForce FX 5950 Ultra. The graphics card based on GeForce FX 5900 XT, which didn’t manage to jump sky-high even after overclocking is slightly behind GeForce FX 5950 Ultra.
The mainstream GeForce FX 5700 Ultra easily outpaced RADEON 9600 XT and got really close to GeForce FX 5900 XT after successful overclocking.
Enabled full-screen anti-aliasing and anisotropic filtering changed the situation on the diagrams a little bit and aggravated the state of things for the slowest testing participants: RADEON 9800 SE based graphics cards working at nominal frequencies appeared dramatically low. Even after the modifications and overclocking these graphics cards are considerably behind the overclocked GeForce FX 5900/5900 XT, as the latter two solutions boast faster graphics memory.
The graphics engine of this air-simulator uses OpenGL and loads the graphics cards pretty heavily. The gaming settings allow enabling pixel shaders, but this hardly makes any sense. First, enabling pixel shaders reduces the performance dramatically while the water surfaces quality improves not that greatly at all. And second, ATI based graphics cards do not draw the water surfaces correctly in case we enable pixel shaders support. That is why we decided to run the tests for our graphics cards with the maximum graphics quality settings but with disabled pixel shaders. Besides that we also ran the tests for forced FSAA and anisotropic filtering.
The results of our tests appeared limited by the CPU and overall system performance.
ATI RADEON 9800 XT, 9800 PRO, NVIDIA GeForce FX 5950 Ultra and overclocked GeForce FX 5900 and GeForce FX 5900 XT perform almost equally fast, so that it is very hard to single out the leader.
The non-overclocked ATI RADEON 9800 SE based graphics cards lose to everybody but after they undergo a few modifications and get overclocked, their performance speed up quite noticeably. The 128bit model outpaces RADEON 9600 XT and gets to the level of GeForce FX 5700 Ultra, while the 256bit model catches up with GeForce FX 5900/5900 XT.
ATI RADEON 9600 XT leaves behind RADEON 9800 SE and is just a few steps behind NVIDIA GeForce FX 5700 Ultra. However, when we overclock it, its higher overclocking potential helps the solution to catch up with the competitor.
As soon as we enabled full-screen anti-aliasing and anisotropic filtering the graphics cards start receiving heavier workload, and the limitations imposed by the CPU and the system in general do not matter for the graphics cards performance any more.
The test scene is full of woods, which are displayed as a row of textured parallel planes, each texture being a cross-cut of the trees. Therefore, the scene contains a lot of surfaces requiring textures with high level of anisotropy. ATI based graphics cards use a more efficient anisotropic filtering algorithm that is why these cards show better overall results.
ATI RADEON 9800 PRO is just a tiny bit behind RADEON 9800 XT and after overclocking it outperforms the counterpart due to higher core frequency.
RADEON 9800 SE based graphics cards share two last positions when working at their nominal frequencies, but the modification and overclocking bring them to the level of NVIDIA GeForce FX 5900 XT and GeForce FX 5900/5950 Ultra.
ATI RADEON 9600 XT outperforms RADEON 9800 SE with 128bit memory bus and NVIDIA GeForce FX 5700 Ultra, but in 1600x1200 it gives away its leadership because of the failed HyperZ III.
NVIDIA GeForce FX 5900 and GeForce FX 5900 XT have nothing to offer in response to ATI’s top-end solutions, but when we overclocked them, they got as fast as the modified and overclocked RADEON 9800 SE with 256bit bus and are almost neck and neck with the leader of NVIDIA based solutions – NVIDIA GeForce FX 5950 Ultra.
Lock On uses Dierct3D, and unlike the previous air-simulator game it has no problems with enabled pixel shaders. DierctX8 pixel shaders are used to imitate the light refraction effect in the streams of heated air coming from the aircraft engine. Forced FSAA and anisotropic filtering in Lock On doesn’t cause any conflicts as well.
All in all, graphics cards based on ATI chips appeared faster than NVIDIA based solutions.
The scene used in our test is very simple from the geometrical point of view, but it is rich in high-resolution textures, so texturing speed appeared the major factor for the graphics cards performance. So no wonder that RADEON 9800 SE with 128bit memory bus performed almost as fast as RADEON 9800 SE with 256bit memory access, and only in 1600x1200 where the memory bus workload increases, the gap between the two appeared more evident. The same way RADEON 9800 PRO featuring higher chip frequency after overclocking got ahead of RADEON 9800 XT in 1024x768 and 1280x1024.
NVIDIA GeForce FX 5900 and GeForce FX 5900 XT fell notably behind GeForce FX 5950 Ultra, but after overclocking they got closer to the leader of NVIDIA based graphics cards.
With enabled full-screen anti-aliasing and anisotropic filtering we managed to complete the tests only for 1024x768. in higher resolutions the cards with 128MB of onboard graphics memory seemed to be using AGP texturing, as the fps rate at that moment (when we saw the earth surface in the scene) didn’t exceed 1fps.
As soon as the earth disappeared and there remained only the airplane and the sky, the fps rate grew up and the average performance throughout the entire demo record finally turned out quite high. However, this result doesn’t make resolutions beyond 1024x768 suitable for comfortable gaming that is why we didn’t include them into the benchmarks results on the diagrams.
Enabled full-screen anti-aliasing and anisotropic filtering changed the situation slightly in NVIDIA based graphics cards’ favor: there are no numerous surfaces where ATI chips could take advantage of their fast anisotropic filtering algorithm.
It is remarkable that NVIDIA GeForce FX 5900 and 5900 XT based graphics cards couldn’t reach GeForce FX 5950 Ultra even after successful overclocking: maybe the twice as small graphics memory starts telling on the graphics cards performance even in 1024x768 mode.
Command & Conquer Generals: Zero Hour is a 3D real-time strategy (RTS) with excellent graphics for the games of this kind. The gaming graphics engine uses DirectX8 vertex and pixel shaders, though we didn’t see any specific effects resulting from this shader support. At the same time the game loads the graphics cards pretty heavily: there are a lot of building models, soldiers and military vehicles in the scene, there are shadows falling from the clouds and all other objects that can cast shadows, which crawl over the ground, the ground bears the marks left by vehicle wheels and explosions, etc. we ran the tests with maximum graphics quality settings. Besides that, we also tested the cards with forced full-screen anti-aliasing and anisotropic filtering.
In 1024x768 and 1280x960 the graphics cards performance appeared very limited by the processor speed, but in 1600x1200 this influence didn’t work any more.
Nevertheless, the fastest graphics cards based on RADEON 9800 PRO and RADEON 9800 XT performed almost equally fast.
The cards on RADEON 9800 SE as usual occupied the very last positions when tested at their nominal frequencies. But they managed to double the performance after modifications and overclocking. RADEON 9800 SE with 128bit memory bus gets to the level of NVIDIA GeForce FX 5900/5900 XT, and its analogue with 256bit memory bus almost catches up with NVIDIA GeForce FX 5950 Ultra and gets really close to RADEON 9800 PRO.
RADEON 9600 XT outperforms RADEON 9800 SE at nominal frequencies, but after the modification, even the slow 128bit model appears unattainably far ahead.
NVIDIA GeForce FX 5900 is a little behind GeForce FX 5950 Ultra, but successful overclocking makes up for the gap.
GeForce FX 5900 XT based graphics card working at nominal frequencies is just a little ahead of GeForce FX 5900 and outpaces it when working at overclocked frequencies getting closer to GeForce FX 5950 Ultra and overclocked GeForce FX 5900.
The card based on NVIDIA GeForce FX 5700 Ultra managed to defeat only the non-overclocked RADEON 9800 SE. Also it proved faster than RADEON 9600 XT, but only when overclocked.
With enabled full-screen anti-aliasing and anisotropic filtering the situation in low resolutions changed in NVIDIA’s favor. However, in general things remained just about the same: RADEON 9800 PRO, RADEON 9800 XT, NVIDIA GeForce FX 95950 Ultra and overclocked GeForce FX 5900 and 5900 XT lead the race.
Well, let’s sum up the results for our today’s testing participants and try to estimate the benefits of each considered solution from the overclocker’s point of view.
Anyway, I hope this article would be of help to you, guys when buying a graphics card. Good luck!