AGP x8/x4/x2 interface, 128 MB DDR SDRAM in 8 chips on the front and back PCB sides.

Here are some aspects to be pointed out:

1. The MSI's design differs a little from the reference 5900 Ultra card, - the right section that controls the power supply is redesigned. I heard that NVIDIA released a new reference design with a shorter PCB and some components replaced. MSI probably used this design. In general, the card is as long as the GF4 Ti 4600 (NV25):
   
   
   
   
   
   

2. The memory size was cut down from 256MB to 128MB - on the backside there are empty space for missing BGA chips. But the 256bit bus wasn't shortened because the card has 8 32bit memory chips (256 bits in all). The card just doesn't support the dual-bank memory mode.
   
   
   
   
   
   

3. In general, the PCB is sophisticated, - you can see it from the photos.

Now let's have a look at the cooler.

MSI FX5900-VTD128
The cooler consists of the bigger and smaller parts: the former is used for the GPU and the memory chips above, while the latter is attached to the backside against the place where the processor is mounted. The bigger cooler, which is extolled above is made of copper and well adjusted for the card's configuration. There are 6 clips to hold it there! The small cooler is rather a victim of fashion than a real necessity. What's the use of a cooler which is pressed to the backside with a rigid thermal element when there is a heap of logic elements between this cooler and the textolite? It's almost useless. Also, if you look at the last photo in this table, you will see that this piece of decoration can make problems if you try to insert such cards into mainboards where the north bridge is turned at 45 degrees and placed too close to the AGP slot. For example, almost all i875P based boards have such configuration. You have to turn back the north bridge's heatsink's pins to let the small cooler of the video card get through. And what if a north bridge has a cooler with a fan and decorative grills? Howeve, you can take off this small cooler, that is why this problem is not critical.

as we can see, this is the production card's marking; the chip was produced on the 19th week of 2003 (approx. the beginning of May). The package is FlipChip, and the die is covered above with a protective lid. A while ago we took off the cover from the NV30 (FX 5800), and can show you how it looks inside:

The card is equipped with the Philips 7114 codec to control the VIVO:

The box contents:

The card ships in the retail package.

The software suite that MSI ships with its cards was described in the GeForce FX 5600 based card review.

Testbed and drivers

Testbed:

• Pentium 4 3200 MHz based computer:
• Intel Pentium 4 3200 MHz CPU;
• DFI LANParty Pro875 (i875P) mainboard;
• 1024 MB DDR SDRAM;
• Seagate Barracuda IV 40GB HDD;
• Windows XP SP1; DirectX 9.0a;
• ViewSonic P810 (21") and ViewSonic P817 (21") monitors.
• NVIDIA drivers v44.65/44.03.

VSync off, S3TC off in applications.

Cards used for comparison:

• Reference card NVIDIA GeForce FX 5900 Ultra (450/425 (850) MHz, 256 MB);
• Sapphire Atlantis RADEON 9800 PRO (380/350 (700) MHz, 256 MB, driver 6.343/6.360);
• Hercules 3D Prophet 9800 PRO (RADEON 9800 PRO, 380/340 (680) MHz, 128 MB, driver 6.343/6.360).

Test results

Before we start examining 2D quality, I should say there are no complete techniques for objective 2D quality estimation because:

1. 2D quality much depends on certain samples for almost all modern 3D accelerators;
2. Besides videocards, 2D quality depends on monitors and cables;
3. Moreover, certain monitors might not work properly with certain video cards.

With the ViewSonic P817 monitor and BNC Bargo cable the card showed excellent quality at the following resolutions and clock speeds:

Test results: performance in 3DMark03

Conventional signs: ANISO 8xP - Anisotropic 8x Performance (earlier it was called Balanced), ANISO 8xQ - Anisotropic 8x Quality, ANISO 16xQ - Anisotropic 16x Quality.

The other day Aleksei Nikolaichuk published the article devoted to cheats in the 3DMark. He carefully described what the programmers at NVIDIA and ATI do with 3DMark2001, in particular, shader replacement. The problem is deeply studied. In short, today all testers actually measure a degree of smartness of the guys at ATI and NVIDIA as well as of FutureMark instead of estimating cards' performance in the 3DMark. FM turned out to be a prostitute. If it's so simple to replace shaders and the developers keep silence, what can we call it yet? That concerns 3DMark2001. Strangely enough but they started protesting against cheats only in the 3DMark03. At the same time it turned out that NVIDIA wasn't a beta tester anymore as the company didn't want to pay through the nose every year. The guys at NVIDIA were denounced by the world community, and FutureMark released a patch where all NVIDIA's tricks were allegedly done away with. ATI was also blamed a little. It reminds me Mark Twain's Tom Sawyer when Tom's Aunt suspected him of all faults, while his brother was just a little reproved. NVIDIA was getting more violent and tougher regarding the 3DMark03.

Suddenly... Aunt made it up with nasty Tom, and Tom even bribed Aunt (who became the beta tester of the future 3DMark), and the latter forgave all sins and faults and called it just innocent pranks. It's clear that Tom will keep on playing his tricks.

So, NVIDIA and FutureMark signed a peaceful agreement where they defined all earlier found cheats as optimizations. Now comes the next driver version 44.65. Let's see whether the guys at NVIDIA have overcome all obstacles of the latest patch 3.30. First of all, let's estimate the speed.

3DMark03, Game1

When we tested the performance with the patch 3.30 that disarmed all NVIDIA's tricks, the speed fell down by 20-24%. Now, the drivers 44.65 bring the speed boost by almost the same percentage.

3DMark03, Game2

There is also a little performance gain with the drivers 44.65 relative to the 44.03.

3DMark03, Game3

It's just the same.

3DMark03, Game4

Wow! The speed with the v44.65 has immensely risen. The gain is even greater than the gap between the v3.20 and v3.30. If you remember, in this test the water surface quality got worse (because of the low accuracy of calculations).

3DMark03, 3D MARKS

So, the overall performance with the patch 3.30 and drivers v44.65 got higher.

3DMark03, Vertex Shader 1.1

The optimizations didn't omit this test either.

3DMark03, Pixel Shader 2.0

Hm.. a while ago we could see the same, even the proportions were the same. The transition from v42.* to v43.51 was followed by a good jump in the pixel shader speed. Since no other tests with the PS load didn't have such results, that was an obvious optimization. I think it's the same here.

The FX 5900/5900Ultra demonstrates a good speed as compared to the others. It just falls behind a little in the shader tests, but I do not doubt that NVIDIA will find a smart programmer for every slow shader! NVIDIA employs so many of them! ;-)

The drivers v44.65 works wonders: the FX 5900 (non-Ultra!) outsmarts its more expensive competitor RADEON 9800 PRO 128MB. But isn't the image quality the cost again? Besides, as Aleksei Nikolaichuk says in his article, cheats change the effect, and the benchmark doesn't operate as it was originally planned by the benchmark makers. That is why we measure a degree of smartness of programmers and driver developers rather than a real card's speed.

Where's the truth? The patch 3.30 plus the driver 44.03? Or this tandem has also suffered from FutureMark's revenge as well? Or maybe we should let them make optimizations on condition that the quality doesn't get worse? What's the truth? The answer is not simple. I will try to answer it partially today, but more about it see in our next review where the driver will be deprived of all possible cheats.

Now comes the quality issue.

ATTENTION! The full-sized screenshots are recorded in the BMP format! EACH SCREENSHOT IS 5.5 MB and packed into a RAR archive of the size varying from 1.7 to 3.2 MB!

Test results: quality in 3DMark03 v.3.30, Game4

As we mentioned above, in the Game4 of the 3DMark03 3.20 the FX5xxx family used the lower accuracy of calculations to draw the water surface, and the picture was deformed (that was shown in this review). The patch 3.30 returns the previous quality level. The drivers 44.65 do not cause such quality degradation anymore. But it's obvious that the shaders were changed again. You can see it on the animated GIF and when comparing the BMP. There are definitely some changes. But in which direction? The question is not simple. We used the reference pictures for comparison though we understand that every accelerator uses its own approaches to form scenes; but anyway, it's interesting to see what's the difference. The picture on the drivers v44.03 is farther from the reference one than that obtained with the drivers v44.65. It turned out that the version 44.65 brought the FX5900 closer to the reference, right? Well, the comparison proves that. But does this comparison indicate cheats, and is FutureMark right saying that the patch 3.30 and the driver 44.03 go the right way? - Nobody knows that yet.

Unfortunately, we couldn't get the 1500th frame with the reference rasterizer (because of too lengthy calculations which ended with errors at that).

Let's use the AntiDetect to make the driver unable to detect applications, and compare the picture the GeForce FX 5900 renders with the "unbiased" driver.

The pictures do differ from the reference one, that is why the picture obtained with the software rasterizer couldn't be considered reference. But why does the picture on the drivers v44.03 and patch 3.30 also differ from the one rendered with the AntiDetect? Not all cheats are removed? Or FutureMark brought in there new ones for the driver detection to artificially decrease the speed? The fact is that the benchmark developers lie saying that the patch 3.30 works correctly.

Test results: quality in 3DMark03 v.3.30, Game1

Again, NVIDIA's drivers use some cheats. This time it's connected not with shaders, but with forced compression of semitransparent textures which are used to form puffs of smoke. This explains the speed boost that was seen above. The pictures of both companies differ from the reference one and from each other. Where is FutureMark now which a while go blamed cheaters so angrily and released the patch 3.30 where all things allegedly fell back into place? We don't see whether the reference pictures are considered correct. Or maybe these are not cheats but useful optimizations? But a picture must be rendered as the developers planned, mustn't it? Where's the truth?

That is the question! Some people in the forums advocate pure, unbiased drivers. But will that be true? I doubt it. However they will be satisfied having fooled NVIDIA and ATI, but they forget that programmers at FM are not less cunning.

Let's make again the driver unable to detect applications with the AntiDetect, and look at the picture that the GeForce FX 5900 renders with such unbiased driver.

There is again some difference from the reference picture, - why does the picture rendered with the drivers 44.03 and patch 3.30 differ from that rendered with the AntiDetect? Do the authors of the 3DMark03 keep their promise?

3DMark03 v.3.30, Vertex Shader 1.1

Taking into account the speed gain of the FX 5900 on the drivers v44.65, it's obvious that the shaders are replaced.

Again we use AntiDetect to make the driver unable to detect applications, and then look at the picture that the GeForce FX 5900 renders with such unbiased driver.

The question remains the same: why does the picture rendered with the drivers 44.03 and patch 3.30 differ from that rendered with the AntiDetect?

3DMark03 v.3.30, Pixel Shader 2.0

It's obvious that the GF's pictures are different on the 44.03 and on the 44.65. The BMP comparison clearly shows that. There is no difference between ATI's drivers. Why? Because mathematics behind methods in different driver versions are the same, they produce the same final picture.

Let's make the driver again unable to detect applications with the AntiDetect and look at the picture that the GeForce FX 5900 renders with such unbiased driver.