NVIDIA GeForce FX 5950 Ultra, GeForce FX 5700 Ultra, GeForce FX 5700, GeForce FX 5600XT

Video cards features
Testbed, test tools, 2D quality
Test results: Quake3 ARENA (speed & quality)
Test results : Serious Sam: The Second Encounter (speed & quality )
Test results : Return to Castle Wolfenstein (speed & quality )
Test results : Code Creatures DEMO (speed & quality )
Test results : Unreal Tournament 2003 (speed & quality )
Test results : Unreal II: The Awakening (speed & quality )
Test results : RightMark 3D (speed & quality )
Test results : TRAOD: Benchmarking in paris5_4
Test results : TRAOD: Benchmarking in paris1c
Test results : TRAOD: Benchmarking in paris2g
Test results : HALO: Combat Evolved (speed & quality )
Test results : Half-Life2(beta): ixbt07 benchmark (speed & quality )
Test results : Half-Life2(beta): coast benchmark (speed & quality )
Conclusion

Following ATI with its RADEON 9800 XT and 9600 XT, NVIDIA announces a new line of products. If you think that they simply overclocked the old models, you will be partially right. For the middle sector NVIDIA produces not the overclocked NV31 but a redesigned processor codenamed NV36.

First of all let's look at the new line:

NVIDIA GeForce FX 5950 Ultra - former NV38, GeForce FX 5900 Ultra overclocked up to 475/475 (950) MHz (recommended price of $399 or 499: question is opened), positioned against ATI RADEON 9800 XT;
NVIDIA GeForce FX 5700 Ultra - former NV36, it's a half of the NV38: 4 rendering pipelines, 4 texture units, 3 vertex pipelines, 475/450 (900) MHz (recommended price of $199), positioned against ATI RADEON 9600 XT;
NVIDIA GeForce FX 5700 - the same NV36, but the clock speeds are lower - 425/275 (550) MHz (recommended price of $149), positioned against ATI RADEON 9600 or 9600 PRO;
NVIDIA GeForce FX 5600XT 128bit - the old NV31 (FX 5600) based on simpler PCB with reduced clock speeds of 230/200 (400) MHz (recommended price of $99), positioned against ATI RADEON 9600 SE. 64bit cards are also expected (at a lower price).

So, there are three new video cards, the 5600XT is just a slower version of the FX5600 meant to press the FX5200 down to the lower-price niche and competes against such ATI's products as RADEON 9600 SE. However, much will depend on the price. It's a bit difficult to oppose the cards because, for example, the recommended price for the FX 5700 is between the RADEON 9600 and RADEON 9600 PRO. The RADEON 9600 SE may also fall below the $80 mark bringing the RADEON 9600 close to $100.

That is why we oppose the cards the way they are approximately positioned on our local market. The thorough comparison can be found in our future 3Digest.

We had only TWO video cards in our lab: 5950 Ultra and 5700 Ultra. The speed results of the FX 5700 are obtained by reducing the clock speeds, like the scores of the FX 5600XT 128bit (slower FX 5600). That is why these figures are approximate, though NVIDIA's products seldom differ in speed if the cards of the same architecture and different default clock speeds are tested on the same frequency.

As to the NV38, it makes no sense to examine its architecture half a year after the release of the NV35. See the details in this review. Here are its brief specs:

Specification

0.13 micron process, copper connections
135 M transistors
3 geometry processors (exceed the specs of the DX9 VS 2.0)
4 enhanced pipelined pixel processors. Instead of the expected 8 pixel processors NVIDIA leaves the same 4 processors. Each of them is equipped with two texture filtering units, two integer-valued floating-point ALUs and one simply floating-point ALU. It results to three floatong-point operations. Such configuration allows up to 12 pixel operations per clock.
AGP 3.0 (8x) system interface
256-bit DDR1 interface (!)
Intellisample HCT (High Compression Technology) is the extension of the Intellisample in NV30 (local memory optimization technology).
Tile operations: caching, compression, Early HSR, Early z Cull
UltraShadow technology - speeds up rendering of stencil shadows (they are called stencil shadows because the stencil buffer if used much when they are generated) at the expense of limitation of the scene depth for objects processed.
Support for high-precision integer-valued formats (10/16 bits per components) and high-precision floating-point formats (16 and 32 bits per components) for texture and frame buffers.
Through accuracy of all operations - 32-bit floating-point arithmetic (128-bit color depth supported)
New optimized anisotropy algorithm, when enabled by user, reduces the performance drop (fps) without quality degradation
Anisotropy quality up to 8x of bilinear filtering, i.e. up to 32 discrete texture samples per one texture fetch
AA hybrid modes - 8x and 6xS
Frame buffer compression allows reducing the performance drop with FSAA enabled
Two integrated 400 MHz RAMDACs
Integrated interface for external TV-Out chip
Three integrated TDMS channels for external interface DVI chips

The NV38 is the same NV35 which meets stricter requirements. The engineers at NVIDIA and TSMC had half a year to improve the production, and now at the output equal to the 5900Ultra they get chips working stably at 475 MHz instead of 450 MHz. The card's also equipped with faster memory. It's quite possible that the GeForce FX 5900 Ultra will be taken out of production to give way to the 5950 Ultra.

The NV36 is a new solution. Have a look at the comparison table:

	NV36	NV31	NV34
Technology, nm	130	130	150
Transistors, M	85	75	47
Pixel pipelines	2/4(1)	2/4(1)	2
Vertex pipelines	3	2	1
Texture units	4	4	4
Core clock, MHz	475 (Ultra)	400 (Ultra)	250
Memory bus, bits	128 (DDR II)	128 (DDR) (2)	128 (DDR)
Memory bus clock (eff.) MHz	900 (Ultra)	800 (Ultra)	400
Pixel shaders	2.0+	2.0+	2.0+ (3)
Vertex shaders	2.0+	2.0+	2.0+
Memory bandwidth, GB/s	14.4	12.8	6.4
HSR	Yes	Yes	Yes
Early Z test	Yes	Yes	Yes
Z compression	Yes	Yes	Yes
Color compression in MSAA modes	up to 1:4	up to 1:4	No
Hardware geometry unit	yes	yes (4)	yes (4)
RAMDAC, MHz	2x400	2x400	2x350
TV-out	integrated	external	external
DVI	integrated	integrated	integrated
Package	FCPGA	BGA (5)	BGA (5)
External power supply	obligatory	desired (Ultra)	optional

Note:

NV31 supports 4x1 (pipelines x texture units) or 2x2.
DDR2 supported
There's some difference between pixel shaders of NV34 and NV31, but both comply with the 2.0 standard.
The hardware geometry processors of the NV31 and NV34 have the same (!) computational performance, twice lower than that of the NV30, and in the NV36 the performance is higher at the expense of three pipelines.
NV25 compatible

The NV36 looks stronger than the NV31. The tests will show how it helps. Note that the GeForce FX 5700 Ultra card is equipped with the DDR-2 (!) memory and its recommended price is only $199! Either the price for the DDR2 fell down much or NVIDIA found a new way to cut the cost price. But this is a question of the design.

Cards

NVIDIA GeForce FX 5950 Ultra
NVIDIA GeForce FX 5700 Ultra

The cards have AGP x8/x4/x2 interface, the FX5700 Ultra has 128 MB DDR-2 SDRAM memory in 8 chips on both PCB sides, FX5950 Ultra has 256 MB DDR SDRAM in 16 chips on both sides.
NVIDIA GeForce FX 5950 Ultra
Hynix 2ns memory chips (corresponds to 500 (1000) MHz), memory clocked at 475 (950) MHz, GPU at 475 MHz. 256 bit memory bus.
NVIDIA GeForce FX 5700 Ultra
Samsung (GDDR2) 2.2ns memory chips (corresponds to 450 (900) MHz), memory clocked at the same frequency, GPU at 475 MHz. 128 bit memory bus.

Comparison with the reference design, front view
NVIDIA GeForce FX 5950 Ultra	Reference card NVIDIA GeForce FX 5900 Ultra


NVIDIA GeForce FX 5700 Ultra	Reference card NVIDIA GeForce FX 5600 Ultra

Comparison with the reference design, back view
NVIDIA GeForce FX 5950 Ultra	Reference card NVIDIA GeForce FX 5900 Ultra


NVIDIA GeForce FX 5700 Ultra	Reference card NVIDIA GeForce FX 5600 Ultra

Here's the reference GeForce FX 5700 based card:

The card is based on the design close to the FX 5600, and judging by its specs it will have the 3.6ns memory in TSOP package.

The design of the FX5950 Ultra is slightly altered compared to the FX 5900 Ultra, in particular, where the power supply unit is placed. The card looks bulky because of the cooler.

Have a look at two flagships of today's 3D graphics sector (RADEON 9800 XT above, GeForce FX 5950 Ultra below):

The cooler reminds the FlowFX - the cooler that spoiled the reputation of the NV30. The first PCI slot is not available...

But it's a different cooler:

Air is drawn from inside of the PC case, at the backside of the heatsink, i.e. the air flow does not change its direction, making the cooler quieter compared to the FlowFX. But it's still noisy as the speed of rotation is pretty high and air goes through a turbine.

If you are interested in the cooler's design look at how we assembled it:

And now comes the NV38 GPU itself:

The chip was manufactured at 27th week, 2003, i.e. in the middle of the summer. Also you can see that figures "5950" are simply added on top. It proves that it is the same GeForce FX 5900 Ultra which just reliably operates at 475 MHz.

The card's equipped with the Philips 7108 coder controlling VIVO:

Quality of such multimedia functions were examined in the previous reviews.

The NV36 design looks more interesting. The PCB of the GeForce FX 5700 Ultra is unique and has only several traits common with the GeForce FX 5600 Ultra. The card has GDDR2 memory instead of the traditional DDR. Such chips, only faster samples, were also installed on the GeForce FX 5800/5800 Ultra.

That PCB was very complicated and NVIDIA had to place orders for cards with third manufacturers, and its partners actually received finished solutions. The current situation is simpler but the PCB is still complicated as the GDDR2 required good protection from pickups:

It's clear that the memory section is entirely screened on both PCB sides. In fact, the card looks very similar to the NV30: GRRD2, 128bit bus, PCB screening, clock speeds... everything but the price. This time the recommended price is $199. I do not believe that all the components got twice cheaper... Either those cards weren't so expensive (NVIDIA just tried to hide it), or this time NVIDIA knowingly incurs losses.

I don't understand why they use GDDR2 memory, which is more expensive and requires dearer PCB design, for the clock speeds of 450 (900) MHz if there's 2.2ns memory of the DDR standard (the GeForce FX 5900 uses exactly this memory type). Is GDDR2 cheaper than DDR? - I don't think so.

The cooler is pretty compact, and the card takes only one slot:

The cooler consists of two parts located on both side.

The cooler mount is reliable, and the cooler itself works quietly.

Here's the chip.

It's the first time we see the FCPGA packed processor with the cap removed made by NVIDIA. I tried to remove the cap from the GeForce FX 5600 Ultra and the chip easily reached 500 MHz (at the default clock speed of 400 MHz). Probably, the guys at NVIDIA arrived at the same conclusion that it's better to cool the die directly.

Also note that the NV36 is produced at IBM's plants, and NVIDIA says that the first revision works pretty well.

In the overclocking mode the NV38 reached 520/1050 MHz and the NV36 demonstrated the impressive jump to 600/1100 MHz! It means that the chip has an excellent potential (the percentage of valid chips for 475 MHz is probably high), and the PCB is competently designed as the GDDR2 memory managed to speed up by 100 MHz in spite of its high temperature. The cards were equipped with additiional fans.

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 52.16 (WHQL!).

VSync off, S3TC off in applications.

The drivers v5x.xx provide the following panels of settings:

The new version gives you flexible tools to manage the AA and anisotropy (you can let the application decide to enable or disable them, or do it yourselves), has Quincunx AA mode changed into 2xQ. Besides, the clock speeds at different resolutions now change simultaneously for D3D and OpenGL. The new version also offers finer adjustment of the clock speeds and resolutions in 2D. In the menu editor you can hide unnecessary menu options. The other features are similar to the versions 4x.xx.

Test results

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

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

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

NVIDIA GeForce FX 5950 Ultra	1600x1200x85Hz, 1280x1024x120Hz, 1024x768x160Hz
NVIDIA GeForce FX 5700 Ultra	1600x1200x85Hz, 1280x1024x120Hz, 1024x768x120Hz

Test results: performance

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

Some time ago we decided not to compare ATI's maximum anisotropic quality of 16x to two NVIDIA's modes anymore. The ANISO 8x Quality mode provided the real maximum quality with both trilinear filtering and anisotropy working to their full capacity. The ATI 16x Quality showed sharper images due to the 16th degree but on some surfaces the filtering quality was lower. That's the way ATI's anisotropy works. That is why we thought it was more correct to compare this ATI's mode with NVIDIA's Performance and Quality.

But NVIDIA's optimization policy changes the things and we do not know anymore if there are applications where NVIDIA's anisotropy works to its full capacity. That is why we consider that it's correct to compare ANISO 16xQ (ATI) to ANISO 8xQ (NV). Both have their strong and weak points, but in general they compensate each other.

Also remember that we compare visual quality, i.e. what we really see. There's no per-pixel comparison as you won't notice difference in quality if it touches just several pixels.

Test applications:

Return to Castle Wolfenstein (MultiPlayer) (id Software/Activision) - OpenGL, multitexturing, ixbt0703-demo, test settings - maximum, S3TC OFF, the configurations can be downloaded from here

Serious Sam: The Second Encounter v.1.05 (Croteam/GodGames) - OpenGL, multitexturing, ixbt0703 demo, test settings: quality, S3TC OFF

Quake3 Arena v.1.17 (id Software/Activision) - OpenGL, multitexturing, ixbt0703 demo, test settings - maximum: detailing level - High, texture detailing level - #4, S3TC OFF, smoothness of curves is much increased through variables r_subdivisions "1" and r_lodCurveError "30000" (at default r_lodCurveError is 250 !), the configurations can be downloaded from here

Unreal Tournament 2003 v.2225 (Digital Extreme/Epic Games) - Direct3D, Vertex Shaders, Hardware T&L, Dot3, cube texturing, default quality

Code Creatures Benchmark Pro (CodeCult) - the game that demonstrates card's operation in DirectX 8.1, Shaders, HW T&L.

Unreal II: The Awakening (Legend Ent./Epic Games) - Direct3D, Vertex Shaders, Hardware T&L, Dot3, cube texturing, default quality

RightMark 3D v.0.4 (one of the test scenes) - DirectX 8.1, Dot3, cube texturing, shadow buffers, vertex and pixel shaders (1.1, 1.4).

Tomb Raider: Angel of Darkness v.49 (Core Design/Eldos Software) - DirectX 9.0, three demo scenes, maximum quality, Depth of Fields PS20 off. The settings are equal for all the cards tested.

If you need patch 49 which is not easy to find and the demo benchmarks let me know by email.

HALO: Combat Evolved (Microsoft) - Direct3D, Vertex/Pixel Shaders 1.1/2.0, Hardware T&L, high quality

Half-Life2 (Valve/Sierra) - DirectX 9.0, two different demo (ixbt07 and coast). Tested in anisotropic filtering mode and in the heavy mode with AA and anisotropy enabled.

Attention! Since this is the beta version that leaked into the Net, we won't take into account the cards' scores obtained in this test.

Note that the results of the GeForce FX 5700 are obtained by reducing the clock speeds of the 5700 Ultra (though the cards will have memory of the different architecture); and the 5600XT was tested as FX 5600 running at 230/400 MHz.

Quake3 Arena

Light modes without AA and anisotropy: the FX 5950 Ultra easily beats its competitor from ATI's camp. The FX5700 Ultra, FX5700, and FX5600XT come out winners.

AA enabled: the FX5950 Ultra still wins in 1024x768 but in higher resolutions the performance gain decreases and in 1600x1200 the card loses. The FX5700 Ultra wins in all resolutions, so does the FX5700.

Anisotropy enabled: NV38 has a good breakaway, and the success of the NV36 is weightier.

AA & anisotropy enabled: the NV38 wins only in 1024x768; the higher the resolution, the stronger the RADEON 9800 XT. The NV36 (both cards) holds the crown reliably.

So:

NVIDIA GeForce FX 5950 Ultra - a complicated situation, though it has a small advantage
NVIDIA GeForce FX 5700 Ultra - victory!
NVIDIA GeForce FX 5700 - victory!
NVIDIA GeForce FX 5600XT 128bit - victory!

RADEON 9800, driver 6.387	FX 5900, driver 52.xx
No AA, No ANISO

AA4x, ANISO 16xQ/8xQ

No complaints about the quality. The FX 5950 doesn't differ from the 5900, all the FX 5700/5600 cards demonstrate the same quality.

Serious Sam: The Second Encounter

Light modes without AA and anisotropy: the NV38 and NV36 outscore their competitors by a little margin. The FX5600XT comes out a winner.

AA enabled: the NV38 as well as the FX5700 Ultra, strengthen their positions. The FX5700 loses to the RADEON 9600.

Anisotropy enabled: the FX5950 Ultra easily wins the battle, and the NV36 (both cards) look more successful.

AA & anisotropy enabled: the NV38 wins! The FX 5700 Ultra performs pretty well, and the FX5700 loses to its competitor.

So:

NVIDIA GeForce FX 5950 Ultra - wins
NVIDIA GeForce FX 5700 Ultra - wins
NVIDIA GeForce FX 5700 - if the FX 5700 is equal to the R9600 PRO in price, it will lose the battle
NVIDIA GeForce FX 5600XT 128bit - leads the pack

RADEON 9800, driver 6.387	FX 5900, driver 52.xx
No AA, No ANISO

AA4x, ANISO 16xQ/8xQ

No complaints about quality (it concerns all FX 5xxx).

Return to Castle Wolfenstein (Multiplayer)

Light modes without AA and anisotropy: the FX 5950 Ultra easily beats its competitor from ATI's camp. The FX5700 Ultra outruns the R9600 XT (though the breakaway in 1600x1200 is not so great), and FX5600XT also comes out a winner. Note that the quality bugs mentioned below are noticed only in FX 5900-5950, not in FX 5600-5700.

AA enabled: NV38 demonstrates no changes, NV36 loses the fight in low resolutions, but in higher resolutions the memory bus helps the FX5700/5700 Ultra to grasp the victory. In the low-level sector the FX 5600 XT is still ahead.

Anisotropy enabled: NV38 loses this time (NVIDIA probably forgets about RtCW from time to time, and some functions work badly: either the speed is low or the quality is not that good). The NV36 also loses in all resolutions except 1600x1200 where it managed to take the lead thanks to its wide memory bus.

AA & anisotropy enabled: the NV38 and NV36 lose the competition in most resolutions.

So:

NVIDIA GeForce FX 5950 Ultra - defeat (complete defeat taking into account artifacts in quality)
NVIDIA GeForce FX 5700 Ultra - defeat (but not by 100%)
NVIDIA GeForce FX 5700 - defeat
NVIDIA GeForce FX 5600XT 128bit - victory

RADEON 9800, driver 6.387	FX 5900, driver 52.xx
No AA, No ANISO

AA4x, ANISO 16xQ/8xQ

The error in NVIDIA's driver is not corrected yet. It appears only in case of the FX 5900-5950.

Code Creatures

Light modes without AA and anisotropy: the NV38 and NV36 outscore their competitors by a little margin. The FX5600XT is more successful.

AA enabled: the NV38, as well as the FX5700 Ultra, strengthen their positions. The FX5700 goes on a par with the RADEON 9600 (if it has the same price as the R9600 PRO, it will lose the competition).

Anisotropy enabled: the NV38 loses the battle, though the FX5950 Ultra still beats the R9600 XT; the FX 5700 looks better.

AA & anisotropy enabled: the FX 5950 Ultra goes on a par with the R9800 XT. The NV36 (both cards) remains a leader.

So:

NVIDIA GeForce FX 5950 Ultra - victory, though the gap is not big
NVIDIA GeForce FX 5700 Ultra - strong victory
NVIDIA GeForce FX 5700 - the same
NVIDIA GeForce FX 5600XT 128bit - perfect!

RADEON 9800, driver 6.387	FX 5900, driver 52.xx
No AA, No ANISO

AA4x, ANISO 16xQ/8xQ

The picture demonstrated by the RADEON and GeForce differ but I can't say that one or the other is better. Tastes differ.

Unreal Tournament 2003

Light modes without AA and anisotropy: the NV38 looks equal to its competitor, the NV36 is a little ahead of its counterpart. The FX5600XT comes out a winner.

AA enabled: the NV38 loses in this test, though both NV36 look good. If the RADEON 9600 PRO is comparable in price to the FX 5700, the latter will have no chances to succeed.

Anisotropy enabled: the FX5950 Ultra demonstrates parity with its competitor. The FX5700 Ultra unexpectedly falls down at 1280x1024, but in general this card wins in this test. The FX5700 easily outdoes its competitor.

AA & anisotropy enabled: the NV38 loses and both NV36 win the competition.

So:

NVIDIA GeForce FX 5950 Ultra - defeat
NVIDIA GeForce FX 5700 Ultra - victory
NVIDIA GeForce FX 5700 - victory
NVIDIA GeForce FX 5600XT 128bit - victory

RADEON 9800, driver 6.387	FX 5900, driver 52.xx
No AA, No ANISO

AA4x, ANISO 16xQ/8xQ

No complains about the quality. The pictures look identical for the whole NV3x family.

Unreal II: The Awakening

No AA, no anisotropy: the NV38 loses, both NV36 are equal to their competitors, the FX 5700 looks more confident. The FX5600 XT loses the battle for the first time.

AA enabled: the NV38 loses here. Both NV36 look equal to their competitors in the popular modes and take the lead in 1600x1200 thanks to their memory bus (but hardly anyone would play in such resolution with AA enabled on such cards).

Anisotropy enabled: the NV38 performs worse. However, all the cards including the NV36 lose this time.

AA & anisotropy enabled: all NVIDIA cards lose in this mode.

So:

NVIDIA GeForce FX 5950 Ultra - defeat
NVIDIA GeForce FX 5700 Ultra - defeat
NVIDIA GeForce FX 5700 - defeat
NVIDIA GeForce FX 5600XT 128bit - defeat

RADEON 9800, driver 6.387	FX 5900, driver 52.xx
No AA, No ANISO

AA4x, ANISO 16xQ/8xQ

No complaints about the quality.

RightMark 3D

No AA, no anisotropy: the whole NV38/NV36 line loses the competition, but shaders were always a problem for the whole NV3x family.

AA enabled: the FX5700 Ultra takes the lead in high resolutions only thanks to its memory bus.

Anisotropy enabled: all the NV36/NV38 cards are beaten.

AA & anisotropy enabled: well, I have nothing to add.

So:

NVIDIA GeForce FX 5950 Ultra - defeat
NVIDIA GeForce FX 5700 Ultra - defeat
NVIDIA GeForce FX 5700 - defeat
NVIDIA GeForce FX 5600XT 128bit - defeat.

RADEON 9800, driver 6.387	FX 5900, driver 52.xx
No AA, No ANISO

AA4x, ANISO 16xQ/8xQ

Look at the bugs in this test with the ATI driver 3.8! The programmers are already informed about it, and we hope that the next time it will be corrected. However, judging by the previous tests where we had no complaints about the quality, the RADEONs look superior. Users do blame the driver v3.8 - how dare they release so raw drivers and in the WHQL pack!

TR:AoD, Paris5_4 DEMO

Shader test... well, all the cards lose here.

So:

NVIDIA GeForce FX 5950 Ultra - defeat
NVIDIA GeForce FX 5700 Ultra - defeat
NVIDIA GeForce FX 5700 - defeat
NVIDIA GeForce FX 5600XT 128bit - defeat.

Quality was already examined, I have no complaints about it.

TR:AoD, Paris1c DEMO

There are fewer shaders, and the NV3x come closer to their competitors, though they are still behind them.

So:

NVIDIA GeForce FX 5950 Ultra - outsider
NVIDIA GeForce FX 5700 Ultra - outsider
NVIDIA GeForce FX 5700 - outsider
NVIDIA GeForce FX 5600XT 128bit - outsider

TR:AoD, Paris2g DEMO

There are almost no shaders, and the FX 5700 grasps the palm. But the NV3x family generally loses this battle.

So:

NVIDIA GeForce FX 5950 Ultra - defeat
NVIDIA GeForce FX 5700 Ultra - defeat
NVIDIA GeForce FX 5700 - victory(!)
NVIDIA GeForce FX 5600XT 128bit - defeat

HALO: Combat Evolved

No AA, no anisotropy: the NV38 goes on a par with the R9800XT; FX5700 Ultra is beaten while the FX5700 takes the lead and outdoes even the R9600 PRO. Unfortunately, the FX5600 XT is smashed to pieces.

Anisotropy enabled: NV38, as well as the FX5700 Ultra, lose this round.

So:

NVIDIA GeForce FX 5950 Ultra - defeat
NVIDIA GeForce FX 5700 Ultra - arguable
NVIDIA GeForce FX 5700 - victory
NVIDIA GeForce FX 5600XT 128bit - defeat

Note that this game doesn't support AA.

RADEON 9800, driver 6.387	FX 5900, driver 52.xx
Example 1
No AA, No ANISO

ANISO 16xQ/8xQ

Example 2
No AA, No ANISO

ANISO 16xQ/8xQ

Example 3
No AA, No ANISO

ANISO 16xQ/8xQ

The quality is good but only in Examples 1 and 2. In Example 3 the GeForce FX hasn't rendered several light sources.

Half-Life2 (beta): ixbt07 demo

I don't include these scores into the overall results as this game version is not finished.

RADEON 9800, driver 6.387	FX 5900, driver 52.xx
Example 1
AA4x, ANISO 16xQ/8xQ

Example 2
AA4x, ANISO 16xQ/8xQ

Example 3
AA4x, ANISO 16xQ/8xQ

Example 4
AA4x, ANISO 16xQ/8xQ

In case of the FX 59xx at the distance of two MIP levels there's blurred strip on the water that removes reflections. There's no such problem with the FX 5600-5700.

Half-Life2 (beta): coast demo

I didn't analyze the results.

RADEON 9800, driver 6.387	FX 5900, driver 52.xx
Example 1
AA4x, ANISO 16xQ/8xQ

Example 2
AA4x, ANISO 16xQ/8xQ

Example 3
AA4x, ANISO 16xQ/8xQ

Example 4
AA4x, ANISO 16xQ/8xQ

Conclusion

So, we have tested the new solutions from NVIDIA, though most of them are just overclocked versions of the old GPUs. But the NV36 is a fully redesigned chip.

NVIDIA is mostly aimed at the wide penetration of the mid-level market with its GeForce FX 5700/5700 Ultra cards, while the NV38 was released just to improve the company's image.

NVIDIA GeForce FX 5950 Ultra is the most powerful 3D solution from NVIDIA. Scores:

Quake3 Arena v.1.17 - wins
Serious Sam: The Second Encounter v.1.07 - wins
Return to Castle Wolfenstein (MultiPlayer) - defeat
Code Creatures Benchmark Pro - wins
Unreal Tournament 2003 v.2225 - defeat
Unreal II: The Awakening - defeat
RightMark 3D v.0.4 - defeat
Tomb Raider: Angel of Darkness v.49 - defeat
HALO: Combat Evolved (Microsoft) - defeat

Summary: this card is approximately equal to its competitor taking into account that shader games are not widely popular yet. But if price of NV38 will be $399 (not $499) and if the price gap remains the same, the NV38 will beat the other as all advantages of the R9800XT in the shader speed are not worth paying $100. I repeat: only if price will be lower then $499.

NVIDIA GeForce FX 5700 Ultra is an interesting solution. It targets the middle sector having the expensive PCB and GDDR2 memory. Scores:

Quake3 Arena v.1.17 - wins
Serious Sam: The Second Encounter v.1.07 - wins
Return to Castle Wolfenstein (MultiPlayer) - defeat
Code Creatures Benchmark Pro - wins
Unreal Tournament 2003 v.2225 - wins
Unreal II: The Awakening - defeat
RightMark 3D v.0.4 - defeat
Tomb Raider: Angel of Darkness v.49 - defeat
HALO: Combat Evolved (Microsoft) - depends on price

Summary: it generally beats the RADEON 9600 XT. The final verdict will be made when both products reach the store shelves.

NVIDIA GeForce FX 5700 is a mainstream solution. Scores:

Quake3 Arena v.1.17 - wins
Serious Sam: The Second Encounter v.1.07 - equal to the RADEON 9600 (if the price for the RADEON 9600 PRO falls to $140-150, the card will lose the battle)
Return to Castle Wolfenstein (MultiPlayer) - defeat
Code Creatures Benchmark Pro - wins
Unreal Tournament 2003 v.2225 - wins
Unreal II: The Awakening - defeat
RightMark 3D v.0.4 - defeat
Tomb Raider: Angel of Darkness v.49 - defeat
HALO: Combat Evolved (Microsoft) - wins
AquaMark3 (Massive Development) - wins

Summary: it looks quite good compared to the RADEON 9600, but if the prices for the R9600/9600 PRO keep on falling down and the 9600 PRO gets equal to the FX 5700, ATI's solution can become more attractive. Well, let's wait for the real cards and prices. Remember that we estimate its performance on the basis of the FX 5700 Ultra card running at the lower clock speed, and the FX 5700 can have different scores.

NVIDIA GeForce FX 5600XT 128bit is NVIDIA's attempt to bring onto the low-end market a more efficient solution than the GeForce FX 5200. The outcome will also depend on the prices. We compare the card to the RADEON 9600 SE which has a 64bit bus but we also expect FX 5600XT 64bit cards. We will examine such cards as soon as they come onto the scene. Scores:

Quake3 Arena v.1.17 - wins
Serious Sam: The Second Encounter v.1.07 - wins
Return to Castle Wolfenstein (MultiPlayer) - wins
Code Creatures Benchmark Pro - wins
Unreal Tournament 2003 v.2225 - wins
Unreal II: The Awakening - defeat
RightMark 3D v.0.4 - defeat
Tomb Raider: Angel of Darkness v.49 - defeat
HALO: Combat Evolved (Microsoft) - defeat
AquaMark3 (Massive Development) - defeat

First of all, we tested not a real card but the FX 5600 at lower clock speeds, and the scores of the real FX 5600XT 128bit can differ. Secondly, the FX 5600XT 128bit is obviously a leader in no-shader games. But the card was released to make DX9 more popular in the low-end segment. That is why the final verdict will be made when the cards come onto the scene (also remember that we expect 5600XT 64bit cards as well). .

In closing I should say that NVIDIA's new price policy can have a bad effect on the competitor. The GeForce FX 5700 Ultra, which looks also good compared to the RADEON 9600XT, has the equal recommended price of $199. It's only the FX 5700 which at its price of $149 can look equal or a bit inferior to its competitors. Let's wait for the real cards and their prices which are expected in 3-4 weeks.

In our 3Digest you can find full comparison characteristics for video cards of this and other classes.

Andrey Vorobiev (anvakams@ixbt.com)

Write a comment below. No registration needed!