MSI FX5900-VTD128
(NVIDIA GeForce FX 5900) Video Card

and more on the Dishonest Treatment of the 3DMark 2003

CONTENTS

1. Video card's features
2. Testbed configuration, test tools, 2D quality
3. Test results: performance in the 3DMark03, drivers v44.65
4. Test results: changes in the graphics quality in the 3DMark03, Game4
5. Test results: changes in the graphics quality in the 3DMark03, Game1
6. Test results: changes in the graphics quality in the 3DMark03, Vertex Shader
7. Test results: changes in the graphics quality in the 3DMark03, Pixel Shader
8. Conclusion

The new video cards based on the NVIDIA's High-End product - GeForce FX 5900 - have finally reached the market. However, first come not the most powerful FX 5900 Ultra based accelerators but a bit slower ones.

The new family codenamed NV35 includes three cards: 5900 Ultra, 5900 and 5900 Value. What is the last product is not clear yet. The 5900 and the 5900 Ultra have the following memory size and frequencies respectively: 128MB against 256MB, and 400/425 (850) MHz against 450/425 (850) MHz.

In all other respects all three lines are identical, and the 5900 Ultra was earlier carefully studied. This and other reviews of the NV3x based cards are listed below.

Theoretical materials and reviews of video cards which concern functional properties of the GPU NVIDIA GeForce FX

• Analysis of the architecture of NVIDIA NV30 (GeForce FX)
• NVIDIA GeForce FX 5800 Ultra (NV30) - single-page review
• NVIDIA GeForce FX 5800 Ultra (NV30) - multi-page review
• NVIDIA GeForce FX 5600 Ultra (NV31) and GeForce FX 5200 Ultra (NV34) - single-page review
• NVIDIA GeForce FX 5600 Ultra (NV31) and GeForce FX 5200 Ultra (NV34) - multi-page review
• ASUS V9900 Ultra on NVIDIA GeForce FX 5800 Ultra - AA and anisotropy quality
• Gainward FX Powerpack Ultra/1000 Golden Sample and Gainward FX Powerpack Pro/660 TV/DVI on NVIDIA GeForce FX 5800 Ultra and 5200 - scaling (performance vs. CPU clock speed) of GeForce FX 5800 Ultra, performance of GeForce FX 5200
• Leadtek WinFast A300 Ultra MyVIVO on NVIDIA GeForce FX 5800 Ultra - performance of GeForce FX 5800 Ultra vs. CPU clock speed in heavy modes with AA and anisotropy enabled
• MSI FX5800 Ultra-TD8X on NVIDIA GeForce FX 5800 Ultra
• Albatron, Chaintech, Gainward, InnoVision, Leadtek, Palit and Prolink video cards on NVIDIA GeForce FX 5200
• ASUSTeK NVIDIA GeForce FX 5200/5600 video cards
• MSI FX5600-VTDR128 (MS-8912) card on NVIDIA GeForce FX 5600
• Albatron, Leadtek and MSI video cards on the NVIDIA GeForce FX 5200 Ultra
• Prolink PixelView GeForce FX 5600 256MB Golden Limited on NVIDIA GeForce FX 5600
• Gainward FX PowerPack Ultra/760 XP Golden Sample on NVIDIA GeForce FX 5600 Ultra (new revision)
• Leadtek WinFast A310 Ultra MyVIVO on NVIDIA GeForce FX 5600 Ultra (old revision)

We, as a test lab, can simply examine the card and return the verdict judging by the card's capabilities, but it seems that both leading chip makers, ATI and NVIDIA, have no shame and raise the the price for High-End products up to $450-550, though it was $100 less a short time ago. We do understand that memory and PCBs are expensive and it's not cheap to develop chips that go through thick and thin and through numerous improvements, castrations, optimizations etc., but still... are the new accelerators so good that a piece of textolite covered with tears and sweat of workers from the USA, Canada and China is worth so much money?

This is a very complicated and painful question, and the answer won't be simple. But we will try to give it today.

The focus, however, will be on the fruit of collaboration of NVIDIA and FutureMark in the light of the "peaceful" agreement according to which all actions of the programmers at NVIDIA aimed at boosting the GeForce FX performance in the 3DMark03 are considered to be optimizations rather than cheats. We will look at how the FX 5900 works on the drivers' latest beta version 44.65, and compare its speed to the FX 5900 Ultra and RADEON 9800 PRO.

Attention! This review will be followed by a review of a similar card but from ASUSTeK where the FX 5900, 5900 Ultra, and RADEON 9800 PRO will be compared in the game tests.

Now let's return to the card which obviously belongs to the High-End class:

It's impossible to resist the beauty of the sparkling heatsink, I even got up early in the morning to caress the half-sun with the first rays of the real sun :-)).

MSI (Micro-Star International) is well known on the market. This Taiwanese company has a wide range of products from mainboards to optical drives. This is an old and faithful ally of NVIDIA, and it used almost all NVIDIA's processors in its video cards.

MSI is also ahead of many NVIDIA's partners with the supplies of the FX 5900 based cards. The card is also equal in price to the previous solution FX 5800 (non-Ultra).

Card

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 between these two majors. 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, while ATI behaved like a little child who was put in the corner, cried a little and promised not to misbehave anymore.

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), but the comparison of the 500th frame with the reference showed that the FX5900 reveals much more similarity with the reference one than the R9800 does. Is that good or bad? Are deviations from the reference picture possible even without cheats? Too many questions, few answers.

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. However, it's generally not correct. 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.

Besides, the comparison with the reference pictures clearly indicate much difference in case of both ATI and NVIDIA.

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. But note that NVIDIA's most optimized picture is very similar to the one rendered with the RADEON 9800. Compare it with the picture that the developers of the 3DMark03 (remember their rage they exposed in parallel with the release of the patch 3.30) consider to be true (44.03 on the patch 3.30). The difference is noticeable. Does ATI use tricks as well? There is no difference between ATI's drivers. But why didn't Futuremark vent its anger on ATI if the patch 3.30 and the drivers 44.03 render what the packet developers consider to be the reference?

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.

What should be considered a standard? Well, if everyone adheres to the DirectX's ideals, let the reference picture be the sought-for standard (otherwise we will hardly ever stop searching it). Well, none of the tests comply with the standard. But the packet has no reference screenshots with it. But since the reference software rasterizer is provided, the reference pictures obtained can be considered the very standard. However, the reality is different, and every accelerator has its own peculiarities. And again, the driver 44.03 plus the patch 3.30 bring a picture different from the one drawn with the Anti-Detect.

Let's sum up the results of the 3DMark03. It's clear that NVIDIA prefers to camouflage the speed drops in the shader operation using certain tricks that boost the speed in the shader tests. The 3DMark03 is the most popular for today. But ATI is not only keeping pace in using such cheats but also lies by promising to remove them. Moreover, the RADEON 9800 stands farther from the reference than the FX5900 in the Game4.

There is no doubt that the methods that let spend less resources on one or another mathematical function should be welcome. And if the guys at FutureMark do certain things imperfectly, and NVIDIA offers a simple and more optimal way, that is good. But in this case there is another rule that everyone must stick to: a program must work as originally planned (though I wish the developers showed what must be in the end). It's unfair to replace key commands in favor of one or another video card. We thus measure not the video cards performance but smartness of programmers.

At the same time you should remember that an end-user uses a card on the manufacturer's drivers and no one would install drivers deprived of such tricks. The final performance depends on the accelerator's capabilities, drivers and such optimizations.

Soon we will take a look at one more driver deprived of detection of the applications that users run. Aleksei Nikolaichuk has already touched upon it when studying the problem of the 3DMark2001. We will take only the game tests and continue working on the 3DMark03.

If you are strongly against optimizations and welcome only the drivers patched with the AntiDetect, I beg you to think again because it's really useless.

1. The driver can control the accelerator in a more optimal way than a given application. Here is an example from Philipp Gerasimov, developer of the RightMark 3D:

   Realization of the Sin/Cos functions by the 3DMark2001 developers:

   ; c12 vecsin ( 1.0f, -1.0f / 6.0f, 1.0f / 120.0f, -1.0f / 5040.0f )
; c13 veccos ( 1.0f, -1.0f / 2.0f, 1.0f / 24.0f, -1.0f / 720.0f )
; c14 pis( PI, 1.0f / (2.0f * PI), 2.0f * PI, 0 )

; now r1.x = fAngle
; wrap theta to -pi..pi
mul r1.x, r1.x, c14.y ; divide by 2*PI
expp r2.y, r1.x ; get fractional part only
mad r1.x, r2.y, c14.z, -c14.x ; multiply with 2*PI

; compute first 4 values in sin and cos series
mov r2.x, c12.x ; theta^0 ;r2.x = 1
mul r2.y, r1.x, r1.x ; theta^2
mul r1.y, r1.x, r2.y ; theta^3
mul r2.z, r2.y, r2.y ; theta^4
mul r1.z, r1.x, r2.z ; theta^5
mul r2.w, r2.y, r2.z ; theta^6
mul r1.w, r1.x, r2.w ; theta^7

; sin
mul r1, r1, c12
dp4 r1.x, r1, c12.x

; cos
mul r2, r2, c13
dp4 r2.x, r2, c12.x


   14 instructions in all

   
   
   Here are the same functions from NVIDIA:

   ; scalar r0.x = cos(r1.x), r0.y = sin(r1.x)
DEF c0, PI, 1.f/2.f, 2.f*PI, 1.f/(2.f*PI)
DEF c1, 1.0f, -1.f/2.f, 1.f/24.f, -1.f/720.f
DEF c2, 1.0f, -1.f/6.f, 1.f/120.f, -1.f/5040.f
MAD r0.x, r1.x, c0.w, c0.y ; bring argument into -pi, .., +pi range
EXPP r0.y, r0.x
MAD r0.x, r0.y, c0.z, -c0.x
DST r2.xy, r0.x, r0.x ; generate 1, (r0.x)^2, .. (r0.x)^6
MUL r2.z, r2.y, r2.y
MUL r2.w, r2.y, r2.z
MUL r0, r2, r0.x ; generate r0.x, (r0.x)^3, .., (r0.x)^7
DP4 r0.y, r0, c2 ; compute sin(r0.x)
DP4 r0.x, r2, c1 ; compute cos(r0.x)


   9 instructions in all.
   It's obvious that the second way is more optimal.

2. New driver versions will bring new optimizations which will be hidden deeper and more properly. Will one programmer be able to stand against a whole army of guys of NVIDIA and ATI? He will until he's enthusiastic enough. Or maybe he won't at all. And who can guarantee that a new ANTIDETECT version covers all cheats in the drivers? What if some cheats are missed? Can we then affirm that we compare accelerators objectively according to the hardware capabilities? No one has proved that it''s objective.

I think that we have to urge both NVIDIA and ATI to step back from the dirty politics of twisting testers' arms with their cheats, and pay more attention to improvement of the hardware instead of licking their errors with drivers. The testers are actually hostages of lies and pseudo-crowns in the 3D graphics sphere and they have to test the performance of accelerators AS IS. Nothing is offered instead yet. You just must remember that if video card A is faster than video card B, it can partially be on account of the driver developers.

We can't keep silence seeing that the GPU makers use forgery for the sake of profits and resort to all possible means to raise ratings and speeds. We can't but agree that NVIDIA and ATI made a laughing-stock of the testers and turned all reviews into a comedy. We insist that the policy of cheating in the drivers for benchmarks be withdrawn, and we ask both companies to disclose normal specifications of their products instead of showing only what's beneficial for marketing departments.

We also wish FutureMark were less dependent on the fact who are beta-testers among the majors and who are not, and protected its 3DMark better from the outward "invasion". The reputation of the 3DMark03 is damaged anyway.

Conclusion

As to MSI's video card tested, it's clear that this is a perfect product regarding its build quality and accessory pack. The company takes close to its heart production of videocards, especially of the High-End rank. The cooling solution, and the design in general are pretty interesting.

Also remember that the FX 5900 is one of the fastest runners today. It fearlessly competes against the RADEON 9800 PRO having the same or even lower price. Besides, none of the RADEON 9800 PRO based cards is able to beat the accessory suite offered by MSI.

The MSI FX5900-VTD128 can offer you:

1. Excellent speed in 3D (especially with the AA and anisotropy enabled);
2. Technological solution with the DX9 support (such games are right around the corner already);
3. VIVO support;
4. pleasure from the high-quality product equipped with a rich retail package (if one pays $400-420, it shouldn't be just a single card or a card with a CD but an outwardly perfect product).

There is just a small disadvantage: it can be difficult to install such card into some mainboards where a north bridge is located close to the AGP port. There can be not enough place between the AGP and the north bridge's cooler for the smaller fan on the card's backside. However, you can easily remove this fan without detriment to the card.