NVIDIA GeForce FX 5900 and ATI RADEON 9800: ASUS, Chaintech and Gainward against Sapphire

Today we are going to compare the performance NVIDIA GeForce FX 5900 and 5900 Ultra based graphics cards with their rivals from the ATI family: solutions on ATI RADEON 9800 and 9800 Pro. The full set of benchmarks, absolutely new gaming tests, impressive overclocking results and more!

by Tim Tscheblockov
09/11/2003 | 11:08 PM

The today’s performance leaders in the cut-throat competition between ATI and NVIDIA among the gaming graphics chips are undoubtedly R350 and NV35, which were officially announced as ATI RADEON 9800 Pro and NVIDIA GeForce FX 5900 Ultra. Correspondingly, the graphics cards based on these two chips occupy the top of the High-End gaming graphics cards price range. Their pricing starts at $400 per unit. In the most “extreme” cases the prices of these graphics cards may reach the phenomenal sum of $1000, like in case of the graphics accelerator from Gainward, which is equipped with a water cooling system.


However, High-End implies not only GeForce FX 5900 Ultra or RADEON 9800 Pro based graphics solutions. A while after the high-performance NV35 and R350 versions had been announced, NVIDIA and ATI launched their slightly slower modifications of the top graphics solutions. I am talking about NVIDIA GeForce FX 5900 and ATI RADEON 9800. They are based on the same NVIDIA NV35 and ATI R350 chips, but differ from the faster modifications by slightly lower working frequencies and GeForce FX 5900 also features less graphics memory.

To be more exact, the differences look as follows:

Today we are going to take a closer look at a few graphics cards based on NVIDIA GeForce FX 5900 and will compare their performance with that of their No.1 competitor: ATI RADEON 9800 based solutions. Besides, we will try to figure out how big is their lag behind the faster solutions, which we have just mentioned, and will also evaluate their overclocking potential. To make this comparison an exiting read for you, I decided to use not only the traditional benchmarks, but also a few new unusual tests.

Today I will introduce to you three models on NVIDIA GeForce FX 5900 from the well-known graphics card makers – ASUS, Chaintech and Gainward. ATI RADEON 9800 will be represented in our article by a solution from Sapphire.

Closer Look: ASUS V9950

The graphics card from ASUS comes in a big box decorated with an impressive hi-tech collage on the front cover:

The package includes the graphics card, of course, the S-Video-to-RCA converter, the DVI-to-D-Sub converter, additional power supply cable, the user’s manual and an impressive heap of CD-discs with drivers, utilities and games. Among the bundled software pieces, I would definitely like to mention Black Hawk Down and Gun Metal:

The graphics card follows NVIDIA’s original reference design for GeForce FX 5900/5900 Ultra and differs from the “Ultra” version only by the absence of memory chips on the reverse side of the PCB:


The connectors set laid out on ASUS V9950 is pretty standard: D-Sub, DVI-I and S-Video:

Like all graphics cards based on NVIDIA GeForce FX 5900/5900 Ultra, this solution features a special connector for the additional power supply cable. If you power on the system without this power supply cable plugged in, the card will start OK, but will work at lower frequencies.

ASUS V9950 is based on NVIDIA GeForce FX 5900 graphics processor and is equipped with 128MB of graphics DDR SDRAM from Hynix with 2.2ns access time:


The GPU works at 400MHz, while the graphics memory supports 850MHz frequency.

The proper image transfer to the TV-Out is ensured by the GPU integrated unit, however, the digital signal for the monitors connected via the DVI is created and managed by the external DVI transmitter from Silicon Image:

To tell the truth, this graphics card made a twofold impression at first glance. On the one hand, we have pretty nice design; high-quality copper cooler, which covers the GPU and memory chips, has very tight contact with the surface of all elements and produces not too much noise, although it is equipped with two fans. On the other hand, we see a very awkward panel with the company logo, which covers more than half of the area above the cooler fans thus reducing the cooling efficiency and increasing the noise level.

Another unpleasant issue deals with some not very loud but very clear rattle, which appears in games and coincides with the refresh rate. As I managed to figure out, this sound comes from the coils of the graphics core voltage regulator.

However, it turned out that almost all graphics cards on NVIDIA GeForce FX 5900/5900 Ultra discussed in our article can “boast” this annoying drawback.

Closer Look: Chaintech A-FX98

The graphics card from Chaintech is packed into a rather big box decorated with a warrior holding some unknown but scary weapon in his hands:

The package contains an S-Video cable, an RCA cable, a DVI-I-to-D-Sub converter, an additional power supply cable, a user’s manual and a set of CD-discs with all necessary drivers and utilities:

The graphics card repeats NVIDIA’s reference design and differs from the previous solution only by the textolite color and the cooler shape:


Chaintech A-FX98 is equipped with a D-Sub Out, DVI-I Out and a combined Video-In/Video-Out port:

The additional power supply connector has been turned and fastened with its side to the PCB, which is considered to be more safe:

The cooler of Chaintech A-FX98 is designed in the same manner as the cooler, which we have just seen on the graphics card from ASUS: copper footing with glued or soldered thin copper plates. However, in this case there is nothing to hinder proper airflow to and from the cooler fans:

At the heatsink foot you can notice special printed spots, which serve as a contact surface between the heatsink and the cooled elements: the memory chips. For better thermal contact they are also covered with sticky heat conductive pads:

The graphics card from Chaintech is based on NVIDIA GeForce FX 5900 chip (the chip seems to belong to one of the early supplies, because it is marked as NV35). It is equipped with 128MB of graphics DDR SDRAM memory from Hynix with 2.2ns access time:


The nominal working frequencies are pretty standard: 400MHz for the core and 850MHz (425MHz DDR) for the memory.

The TV-Out signal is formed by the GPU integrated unit. The decoding of the TV signal is performed by an external SAA7114 chip from Philips, while the DVI transmitter from Silicon Image is responsible for creation of the digital signal for the DVI-I Out.


Chaintech A-FX98 graphics card made a very good general impression due to the low level of noise and not very high working temperature. The voltage regulator coils of the Chaintech card also generated that rattling noise, however it was considerably quieter than by ASUS.

Closer Look: Gainward FX PowerPack! Model Ultra/1200 XP “Golden Sample”

The package of the Gainward graphics card based on NVIDIA GeForce FX 5900 is designed in a very rare peaceful manner:

The package included a cable serving as a converter from the combined Video-In/Video-Out port to 2 S-Video and 2 RCA connectors. Also there was a DVI-I-to-D-Sub converter, a cable for additional power supply, a user’s manual, Gainward’s brand name shield and a set of CDs with the software:

The card also follows NVIDIA’s referenced design for GeForce FX 5900/5900 Ultra:


Gainward FX PowerPack! Model Ultra/1200 XP “Golden Sample” graphics card is equipped with a D-Sub Out, DVI-I Out and a combined Video-In/Video-Out:

Like all NVIDIA GeForce FX 5900/5900 Ultra based graphics cards, the solution from Gainward requires additional power supply:

The cooling system implemented on this graphics card includes a massive aluminum heatsink covering both: the graphics chip and the memory chips altogether. The heatsink is provided with two fans featuring nice-looking LEDs:

Gainward FX PowerPack! Model Ultra/1200 XP “Golden Sample” is the “brightest” graphics cards of all in our lab. Those of you who like unusual outstanding designs will definitely love it:


The card is based on NVIDIA GeForce FX 5900 chip and features 128MB of DDR SDRAM from Hynix with 2.2ns access time. The GPU used on this particular card was produced even before the official product launch that it why it is marked as NV35:


The nominal working frequencies of the graphics core and memory are 400MHz and 850MHz (425MHz DDR) respectively.

It is very interesting that they used not the traditional DVI-transmitter from Silicon Image to form a DVI-signal, but a similar solution from THine Electronics with higher conversion frequency of 170MHz. The TV signal is encoded by a special unit integrated into the GPU, and the decoding is performed by the popular SAA7108 chip from Philips:


To tell the truth I liked the Gainward FX PowerPack! Model Ultra/1200 XP “Golden Sample” most of all graphics cards described in this article. Although its cooler produces a little bit more noise, the voltage regulator coils start rattling only when the card is overclocked and used in really heavy modes. Besides, the fashionable blue highlighting is very pleasing for the eye.

So, now that we have taken a close look at our today’s testing participants, let’s pass over to the tests!

Testbed and Methods

Our test system was configured as follows:

The graphics cards from ASUS, Chaintech and Gainward based on NVIDIA GeForce FX 5900 follow the reference design and work at the same core and memory frequencies that is why I took only one of them for the tests: it is evident that they will show almost identical results. I decided to use the graphics card from Gainward, as it boasted the highest overclocking potential of all solutions considered (we will discuss it a bit later in the article) and features the most attractive exterior.

Another participant from the NVIDIA’s camp will be a graphics card from ASUS based on NVIDIA GeForce FX 5900 Ultra. The competitors from ATI will be represented by ATI RADEON 9800 and RADEON 9800 Pro from Sapphire. These graphics cards arrived into our testlab in OEM packages. Since they are absolutely identical to ATI’s reference designs for these solutions, it doesn’t maker much sense to describe them in detail.

As a result here is the list of graphics cards, which took part in our test session:

We used the following software:

During the tests we used not only the standard test episodes from Quake3 Arena, Serious Sam: The Second Encounter and Unreal Tournament 2003, but also a few ones I created myself.

The graphics quality settings are represented on the screenshots below:

NVIDIA GeForce FX 5900/5900 Ultra

ATI RADEON 9800/9800 Pro

Please have a look at the anisotropic filtering quality settings. For both: graphics cards on NVIDIA’s chips as well as graphics cards on ATI’s chips we set “quality” anisotropic filtering modes. With a high-end graphics card at hand it would be illogical to compromise, don’t you think so?

Contemporary ATI chips are known for their ability to provide high texturing speed with enabled anisotropic filtering at the expense of slightly lower image clarity at “inconvenient” angles (please, see our ATI RADEON 9700 Pro Graphics Card Review and our NVIDIA GeForce FX 5900 Ultra (NV35) Review). GeForce FX chips and NVIDIA drivers are also optimized for higher performance with enabled anisotropic filtering. However, all these optimizations inevitably cause texture quality worsening, that is why in this test session I decided not to use the modes where these optimizations are activated.

The graphics cards comparison in the Quality mode may seem unfair to you, and it is not for nothing that you get this impression. However, when I decide to test the graphic cards in Quality modes, I will not be able to activate the Performance mode on NVIDIA GeForce FX 5900/5900 Ultra based graphics cards, because in this mode the driver automatically reduces the anisotropy level if possible. Besides, tri-linear filtering sometimes simply vanishes, and instead we can notice very clear texture ripples and borders between the MIP-levels. In other words, enabling the “Performance” mode on NVIDIA GeForce FX 5900/5900 Ultra nearly puts the image quality of NVIDIA chips at the same level with the “Performance” mode of ATI chips. RADEON 9800/9800 Pro however, doesn’t use tri-linear filtering at all and anisotropic filtering is performed really fast: here NVIDIA chips are completely defeated.

As a result, we have to use “Quality” modes and put up with a huge performance drop by NVIDIA GeForce FX 5900/5900 Ultra. Although we should also always keep in mind that NVIDIA chips, unlike the solutions from ATI, provide excellent image quality in any conditions and at any view angles.

Performance: IL-2 Sturmovik: Forgotten Battles

IL-2 Sturmovik: Forgotten Battles is a continuation of the famous Russian flight simulator game called IL-2 Sturmovik, which received a lot of Russian and international awards of the gaming sites and magazines. The excellent graphics quality of the new IL-2 Sturmovik: Forgotten Battles game is provided by a very modern graphics engine which supports adaptive changing of the level of detail for the military equipment and landscapes. Water surfaces in this game look highly realistic due to DirectX 8.0 pixel shaders and the landscapes with picturesque woods look absolutely live due to very unusual drawing techniques applied (we will talk about these techniques later in this article). And look at the sky and clouds, very beautiful!

IL-2 Sturmovik: Forgotten Battles is provided with integrated modules measuring the system performance and a set of preliminarily created test scenes that is why we had no problems testing our graphics cards in this application.

For our tests we used TheBlackDeath.ntrk demo:

When you start the test you should press [Shift]+[Tab] in the very beginning to call a console. The scene will automatically be paused. In the console you should type in fps START SHOW which will call a fps counter in the upper part of the display. If you close the console by pressing [Shift]+[Tab] again, the scene will resume with the fps counter running in the upper part of the screen. In the end of the scene you can stop the test by pressing P and take down the number.

During the tests I used full-screen OpenGL modes by enabling the “Maximum settings” item in the stencil-buffer tab. The image quality in the game settings menu was set to “Ideal”.

The game sets very serious requirements to the graphics cards performance as well as to the performance of the entire system. You can see this from the results obtained in 1024x768: here ATI RADEON 9800 falls just a little bit behind the considerably faster RADEON 9800 Pro, just like NVIDIA GeForce FX 5900 falls a little behind GeForce FX 5900 Ultra. It means that in 1024x768 resolution the results shown by our testing participants are limited not by the fillrate or the graphics memory bandwidth, but by some other factors, which do not depend on the resolution. They are: polygon processing speed, data transfer rate via the AGP bus, or finally the optimization of the OpenGL part of the driver. Anyway, in 1024x768 and 1280x1024 the graphics cards based on NVIDIA chips outperform thei rivals on ATI solutions.

As the resolution grows, the graphics cards get more loaded with work and this is where GeForce FX 5900 and RADEON 9800 start falling behind their faster fellows. I would like to draw your attention to remarkably equal results of the competitors in 1600x1200.

When the full-screen anti-aliasing is enabled, the graphics cards’ performance didn’t suffer too much. I dare conclude that FSAA 4x in IL-2 Sturmovik: Forgotten Battles is no problem for the today’s monsters with 256bit bus and tremendous memory bandwidth. Nevertheless, the situation in higher resolutions changed a little bit in NVIDIA’s favor.

When we enabled anisotropic filtering the picture got completely different. The graphics cards on ATI chips took the lead.

Why so? The reason lies in the peculiarities of the anisotropic filtering algorithms from NVIDIA and ATI, as well as in the gaming engine itself. At first let’s dwell on the gaming engine.

Firstly, to draw the landscape IL-2 Sturmovik: Forgotten Battles seems to be using three texturing layers: two base textures and a “detail” texture reflecting the micro-relief of the surface when you are at the closest distance.

Secondly, a pretty large part of the ground in TheBlackDeath.ntrk demo is covered with woods. The woods are very originally drawn in this game. Instead of drawing each tree separately with the help of sprites or polygonal models, the engine imitates big pieces of woods with the help of five planes located in parallel one above the other at a low level above the ground. The planes are covered with transparent textures with the cross-sections of the trees. Note that these tree cross-sections are located exactly under one another on all five planes. This drawing technique provides excellent visual result at a big angle relative to the horizon even if the distance is not very big. However, as soon as the angle gets less than 30 degrees to the horizon the image turns not so realistic any more. You can even see separate tree cross-cuts.

So, the gaming engine is designed so that landscape drawing requires a lot of texture layers to be applied. Therefore, when anisotropic filtering is enabled it is applied to every texture thus increasing the workload immensely.

And how did our graphics cards perform here? The solutions on ATI chips featuring faster but less efficient implementation of the anisotropic filtering algorithms managed to get far ahead their rivals from NVIDIA.

As soon as FSAA joins the enabled anisotropic filtering the graphics cards performance got somewhat lower, although the overall picture remained the same: ATI based graphics cards were still ahead of NVIDIA based ones although their image quality at “inconvenient” angles was not that impeccable.

Performance: Quake3 Arena

In Quake3 Arena I used the maximum graphics quality settings. For our tests I used a standard four.dm_66 demo:

And also my own demo, which I will mention a little later. So, first let’s consider the graphics cards performance in a standard demo:

NVIDIA GeForce FX 5900 outperforms ATI RADEON 9800 Pro not to mention ATI RADEON 9800. In fact I can’t tell you for sure what helped the leader to be so successful here: the graphics chip, the graphics card or ATI and NVIDIA software developers. There is only one thing we can say with all certainty: Quake3 Arena is definitely among the first on the list of applications, which receive optimizations in the gaming graphics cards drivers.

Judging by the results in 1024x768 where the graphics cards receive the minimal workload and the difference between the “fast” and the “slow” modification of the card is not that great, we can say that NVIDIA’s OpenGL part of the driver boasts better optimizations.

As the resolution grows, graphics cards on NVIDIA GeForce FX 5900/5900 Ultra get farther ahead of the rivals: Quake3 Arena is a very “convenient” application for them. First, most surfaces in the scenes feature two texturing layers (this is optimal for GeForce FX 5900/5900 Ultra, which features four pixel pipelines with two texturing units on each). Second, GeForce FX 5900/5900 Ultra work at higher core and memory clock frequencies and the results in the selected testing modes are primarily determined by the fillrate and memory bus bandwidth.

When we enable full-screen anti-aliasing, graphics cards on ATI chips start falling even farther behind their competitors. For example, in 1600x1200 NVIDIA GeForce FX 5900 is almost 1.5 times faster than ATI RADEON 9800.

The anisotropic filtering algorithms work faster on ATI graphics cards but provide lower image quality. Nevertheless, it allows RADEON 9800/9800 Pro based graphics solutions to catch up with the competitors and even to outpace them in 1024x768 resolution.

In the heaviest modes with forced FSAA and anisotropic filtering NVIDIA solutions manage to regain their leadership. Moreover, NVIDIA GeForce FX 5900 easily defeated the fastest solution from ATI – RADEON 9800 Pro – in 1024x768 and 1600x1200 resolutions.

The tests in the standard four.dm_66 demo indicate an indisputable advantage of the NVIDIA based graphics cards. However, it seems really interesting to find out how big is the software developers’ contribution to this success, and if the Quake3 optimization is just an optimization for a few standard demos and not for the entire gaming engine.

To check this out I created my own demo on q3dm7 map: tim_q3.dm_66:

If you want to create your own demo, you should type \g_synchronousClients 1 into the console. The record should start with the following command: \record demo_name, and end up with \stoprecord.

Well, let’s see how the results of the tested graphics cards will change in the new demo:

The new record is a little bit heavier than the standard one and the performance of the graphics cards is a bit lower here. however, the overall picture remained unchanged: NVIDIA GeForce FX 5900 and GeForce FX 5900 Ultra are ahead of all.

With enabled FSAA NVIDIA’s graphics cards again manage to leave their rivals far behind.

The new record also turned out harder for NVIDIA’s anisotropic filtering algorithms that is why AT RADEON 9800/9800 Pro based graphics cards managed to get a little ahead of their competitors in almost all resolutions.

In the hardest mode of the tim_q3 record, as well as in the standard record, the cards based on GeForce FX 5900/5900 Ultra left ATI RADEON 9800/9800 Pro far behind.

So, the situation with the graphics cards performance has hardly changed as we moved from the standard demo to a completely new record, and it means that the OpenGL part of the graphics cards drivers is optimized for the entire Quake3 Arena engine and not just for the standard records used in most test packages.

Performance: Serious Sam: The Second Encounter

During the tests in Serious Sam: The Second Encounter I used standard 32bit OpenGL modes with Quality graphics settings. I also launched “GFX: Extreme Quality” add-on.

Like in Quake3 Arena, I also used a standard “The Grand Cathedral” demo:

And a new record :)

First come the results of The Grand Cathedral demo:

Serious Sam: The Second Encounter with “GFX: Extreme Quality” settings uses the maximum level of anisotropic filtering, that is 8x for NVIDIA GeForce FX 5900/5900 Ultra and 16x for ATI RADEON 9800/9800 Pro. Although AT chips perform anisotropic filtering faster than their competitors, they lost here.

With enabled FSAA the situation changes even more in NVIDIA’s favor.

Now let’s have a look at the performance of our testing participants in a non-standard record. To create your own demo-record in Serious Sam: The Second Encounter you should press [F7], which will call a blinking word “Recording…”. To stop the record, you should hit [F8] key. The record file with an automatically assigned RecordedXX.dem name will be placed into Temp subdirectory of the game. If you would like to play this demo you should first move it to the Demos directory of the game folder.

I created the tim_sstse.dem on the very first gaming level, Sierra de Chiapas:

So here are the results obtained in a non-standard record:

Very interesting, don’t you think so? Of course, the new record is no heavier than the standard one, however all graphics cards showed lower results in tim_sstse.dem!

Moreover, the whole picture got completely different now: ATI RADEON 9800 outperforms NVIDIA GeForce FX 5900, and RADEON 9800 Pro defeats NVIDIA GeForce FX 5900 Ultra.

With forced FSAA the performance gap between graphics solutions on ATI chips and those on NVIDIA chips got a little smaller, but on the whole the situation remain unchanged: ATI based graphics card prove faster than their competitors from NVIDIA.

I can’t say for sure that there are any real “cheats” for the standard “The Grand Cathedral” demo in ATI’s or NVIDIA’s drivers, since the two records I used differ really greatly from one another. For example, the first record has a lot of monsters in the scene at a time, that is why it also requires fast polygon processing. The second record, on the contrary, has very few monsters, but the biggest part of the workload is most likely to be connected with the grass and leaves drawing, i.e. with the objects featuring only one texturing layer. It is most optimal for the pixel pipelines of ATI RADEON 9800/9800 Pro featuring only texturing unit per pipeline.

Nevertheless, the obtained results allow us to conclude that it would be unfair to call any of the tested graphics cards an indisputable leader in Serious Sam: The Second Encounter taking into account the performance only in one standard demo record.

Performance: Unreal Tournament 2003

During the tests in Unreal Tournament 2003 I used 32bit modes and maximum graphics quality settings: Texture Detail - Highest, World Detail - Highest, Character Detail - Highest, Physics Detail - High, Character Shadows - ON, Dynamic Lighting - ON, Detail Textures - ON, Projectors - ON, Tri-linear Filtering - ON, Decals - ON, Coronas - ON, Decals Stay - High, Foliage - ON, Use Blob Shadows - OFF.

To test the graphics cards I used a standard Antalus Flyby scene:

And my own record, which as usual, will be discussed a little later.

First let’s check the results obtained in a standard demo:

Graphics cards based on NVIDIA chips took the lead in all resolutions except 1024x768, where the laurels were won by RADEON 9800 Pro.

With enabled full-screen anti-aliasing the situation has hardly changed: the leadership still belongs to NVIDIA GeForce FX 5900/5900 Ultra based graphics cards.

Faster but less efficient anisotropic filtering implementation allowed ATI chips to improve the situation a little bit. However, NVIDIA GeForce FX 5900 is still ahead of ATI RADEON 9800, and GeForce FX 5900 Ultra works faster than RADEON 9800 Pro.

Adding FSAA doesn’t affect the results at all.

So, as we have just seen in the standard Antalus Flyby scene, NVIDIA GeForce FX 5900 is faster than ATI RADEON 9800, and GeForce FX 5900 Ultra outperforms RADEON 9800 Pro.

However, Antalus Flyby plays an artificial situation, which has nothing to do with the actual gaming process. In order to really evaluate the graphics cards performance in Unreal Tournament 2003, we should create a record of a real gaming fragment. Luckily, a new patch 2225 for Unreal Tournament 2003 has been released, so that creating records is not a problem any more.

At first you should type demorec record_name in the console, and to stop recording – stopdemo. To test a graphics card you should start the demo record with the following command: demoplay record_name?timedemo.

I created the new record in Botmatch on the gloomiest and most exciting map called DM-Inferno:

Here are the results obtained in the new tim_ut2003.dem record:

Wow, where are those impressive numbers exceeding 200fps? The graphics cards performance dropped 2-3 times down, which is not at all surprising, as the workload grew up tremendously. Now even in 1024x768 the graphics cards performance is not acceptable for comfortable gaming, not to mention higher resolutions.

But how does the whole picture look like now? Well, NVIDIA’s solutions have lost their indisputable leadership. GeForce FX 5900 is only a little faster than ATI RADEON 9800, while the advantage of GeForce FX 5900 Ultra over the rival from ATI is nearly negligible.

With enabled FSAA ATI based graphics cards appeared faster. The graphics cards on NVIDIA GeForce FX 5900 managed to outperform the competitor only in 1024x768, and its elder brother, GeForce FX 5900 Ultra, turned out faster than RADEON 9800 pro only in 1600x1200.

With enabled anisotropic filtering ATI’s chips show better results than their competitors due to faster but less efficient implementation of this function.

Under the heaviest workload NVIDIA GeForce FX 5900/5900 Ultra based graphics cards suffer a complete defeat.

So, the results obtained in the standard Antalus Flyby demo indicated an indisputable victory of NVIDIA GeForce FX 5900/5900 Ultra based graphics solutions. However, in the new record, which was closer to the real gaming conditions, the advantages of NVIDIA’s chips over the rivals from ATI were no longer that evident. And under heavy workload NVIDIA completely lost the leadership.


Graphics cards based on NVIDIA GeForce FX 5900 differ from the Ultra versions by the core clock frequency and the amount of onboard graphics memory. GeForce FX 5900 is equipped with 128MB of graphics memory in eight 32bit chips, which results into a 256bit memory bus. However, GeForce FX 5900 Ultra features 16 chips like that and its memory works in a dual-bank mode. It provides a certain performance boost compared with the 128MB versions. Although efficient overclocking of the available graphics memory can easily surpass this advantage, since the nominal memory frequencies on GeForce FX 5900 and GeForce FX 5900 Ultra are the same: 850MHz (425MHz DDR).

The list below contains the maximum chip and memory frequencies we managed to overclock our testing participants to without any additional cooling system modifications:

As you can see, the graphics cards demonstrated absolutely identical memory overclocking potential: all the three models reached 960MHz frequency. As for the graphics chip overclocking results, the highest frequency was achieved on the card from Gainward, which we actually hadn’t expected at all, at least with an aluminum cooler like that. Well, copper is not everything, as you see now. :) We managed to overclock the graphics core by 27.5% and the graphics memory – by 12.9%. This is a really good result, don’t you agree? The obtained frequencies appeared much higher than those we got during NVIDIA GeForce FX 5900 Ultra overclocking.

So, I decided to run the next bunch of tests on a solution from Gainward (you will certainly get an idea of the others’ results as well, because they will be just a little lower than those of the overclocking leader - Gainward FX PowerPack! Model Ultra/1200 XP “Golden Sample”).

The competitor of NVIDIA GeForce FX 5900, ATI RADEON 9800, also pleased us with its high results:

The graphics chip acquired 44.6% of the nominal frequency (!), and the graphics memory overclocked by 20.7%. As a result, the overclocked RADEON 9800 also appeared faster than its more powerful modification, ATI RADEON 9800 Pro.

Well, the ongoing comparison of the overclocked graphics cards performance promises to be very interesting. Ok, let’s start with Quake3 Arena.

Overclocking: Quake3 Arena

The overclocked NVIDIA GeForce FX 5900 and ATI RADEON 9800 manage to get far ahead of more expensive graphics card models.

Enabled full-screen anti-aliasing requires higher memory bus bandwidth. Even though the memory of our graphics cards overclocked not as successfully as the graphics cores, GeForce FX 5900 and RADEON 9800 retained their leadership.

With enabled anisotropic filtering the graphics processor is loaded much heavier, while the memory bus workload remains almost the same. So, graphics core overclocking is essential particularly for anisotropic filtering. Take a look at the results:

In the hardest mode in the highest resolutions the gap between overclocked graphics cards and the solutions working at the nominal frequencies gets somewhat smaller. It is probably the memory bus workload that starts telling here, especially keeping in mind that memory overclocking was not very efficient.

So, in Quake3 Arena the overclocked NVIDIA GeForce FX 5900 and ARI RADEON 9800 proved faster than their more powerful counterparts. They managed to outpace their non-overclocked Ultra and Pro fellows by the same value as their fellows outpaced them before we sat to overclocking.

Now let’s see how the situation will change in a heavier test: the “gaming” Unreal Tournament 2003 demo record:

Overclocking: Unreal Tournament 2003

In the lightest modes our overclocked graphics cards outperform NVIDIA GeForce FX 5900 Ultra and ATI RADEON 9800 Pro working at the nominal frequencies. However, the performance difference between them is not so great as in Quake3 Arena.

With enabled FSAA the advantages of the overclocked RADEON 9800 over the Pro version gets smaller, and the overclocked NVIDIA GeForce FX 5900 even loses it completely in all resolutions above 1024x768. It is probably bigger amount of graphics memory and its dual-bank working mode that tell here.

As soon as we enable anisotropic filtering, the overclocked graphics cards dash forward. RADEON 9800 appears the fastest in this race, due to faster anisotropic filtering algorithms, which, however, provide not very high image quality.

Under the heaviest workload the overclocked ATI RADEON 9800 is still ahead of the RADEON 9800 Pro. The overclocked NVIDIA GeForce FX 5900 starts falling behind its faster Ultra version in higher resolutions where the memory bus workload grows up.


Well, the performance difference between NVIDIA GeForce FX 5900, ATI RADEON 9800 and their faster analogues appeared not very high: 10-15% in the heaviest work modes. The pricing on NVIDIA GeForce FX 5900 Ultra and ATI RADEON 9800 Pro exceeds that on the non-Ultra and non-Pro versions much more than by these 10-15%.

Therefore, if you are looking not for the “world’s very best” graphics card, but just for a good High-End product with a nice potential to last it for a while, then why don’t you take a closer look at NVIDIA GeForce FX 5900 and ATI RADEON 9800? These solutions become the most interesting today, because if overclocked, they are mostly faster than their more expensive fellows.

The upcoming launch of NVIDIA NV38 and ATI R360, the “overclocked versions” of the NV35 and R350, which is scheduled for this September-October, will not tell on the attractiveness of NVIDIA GeForce FX 5900 and ATI RADEON 9800. On the contrary, the new chips announcement will make the cards based on NVIDIA GeForce FX 5900 and ATI RADEON 9800 less expensive, while the overclocking potential will remain the same :)

It seems to be a more difficult question, which graphics card of the two is better: NVIDIA GeForce FX 5900 or ATI RADEON 9800?

On the one hand, NVIDIA GeForce FX 5900 with a better anisotropic filtering algorithm on the whole performs almost as fast as RADEON 9800 or maybe even a little faster in some cases.

On the other hand, DirectX8 shaders were the maximum our today's benchmarks could really involve, while in case of DirectX9 shaders RADEON 9800 will definitely be much faster. As a result, it will do really great in modern games using DirectX9 shaders. Finally, the price of graphics cards based on ATI chips is considerably lower, than that of NVIDIA based solutions. So, from the pricing point of view, NVIDIA GeForce FX 5900 is more likely to compete not with ATI RADEON 9800, but with a more powerful ATI RADEON 9800 Pro…

Well, undoubtedly you will make the right choice!

P.S.: You can download non-standard record demos, which we used during this test session here (616KB).