NVIDIA GeForce4 Ti 4200 with AGP8x (NV28) and GeForce4 MX 440 with AGP8x (NV18)

Eight x's - what's new?
Peculiarities of NV28 and NV18 based video cards
Test system configuration and driver settings
Test results: briefly on 2D
Test results : 3DMark2001 SE game tests
Test results : Quake3 ARENA
Test results : Serious Sam: The Second Encounter
Test results : Return to Castle Wolfenstein
Test results : Code Creatures DEMO
Test results : Comanche4 DEMO
Test results : Unreal Tournament 2003 DEMO
Test results : AquaMark
Test results : RightMark 3D
Conclusion

Eight x's - what's new?

We know quite well that AGP is a bottleneck in a system which controls data processing in image displaying, namely: CPU bus, memory, north bridge, accelerator, and accelerator's memory. Well, it's one more step towards the perfection. Let's sum up the differences and new features of the AGP 3.0 standard, also known as AGP 8x:

Doubled throughput - 2.1 GB/s, the effective clock speed is equal to 533 MHz:

	AGP 4X	AGP 8X
Bytes per transfer	4 (32 bits)	4 (32 bits)
Clock rate	266.67MHz	533.33MHz
Bus bandwidth	1.1GB/sec.	2.1GB/sec.

The isochronous data transfer (i.e. continuous stream transfer with a guaranteed latency) got new capabilities which let the accelerator know that required parameters will be delivered in time - it makes possible to avoid null cycles.
The bus is entirely (forward and backward) compatible with cards of the previous version (AGP 2.0) and can work in 2x and 4x modes in the 2.0 standard and in 4x and 8x modes in the extended 3.0 standard. So, new cards which comply with the AGP 8x specification will be compatible with motherboards with AGP 2x, 4x and 8x, and new AGP 8x boards will support cards with 2x and 4x provided that they comply with the AGP 2.0 or 3.0 specification. As we know, there are some compatibility problems with earlier implementation of the AGP 8x (SIS 468, R300 etc.).
Optimizations for efficient AGP texturing are promised.
Some rare and unpopular capabilities of the 2.0 will be removed (they can be deleted without losing compatibility).

Well, the paper looks magnificant. But no applications can actually get a significant gain with the AGP 8x (except some specific synthetic tests). I hope such applications will appear soon, but at present AGP 8x is not a weighty argument for a buyer. Well, if one has to pay a lot of money over to get 8x, then it doesn't make sense to get such a card, especially because current bus realizations have some compatibility problems with graphics solutions, for example, from ATI. It's clear that only with the reference implementation of the AGP 3.0 in one of the Intel's chipsets the problems of interpretation of the standard and of compatibility won't be so vital anymore.

It's interesting that the modern accelerators don't use effectively the AGP bus - it's still far even from the theoretical limit of the AGP 4x; only NVIDIA is a step closer to it. However, it is possible that the AGP conception or the way it's realized in the chipsets can also be blamed. I hope the Intel's AGP 8x and NV30 solutions will pump graphics data through the bus much faster.

Without further ado let's turn to the cards we are testing today.

Cards

The cards have AGP x2/x4/x8 interface, 128 MB DDR SDRAM (8 chips on both PCB sides) on the GeForce4 Ti 4200 with AGP8x (NV28), and 64 MB DDR SDRAM (4 chips on the front PCB side) on the GeForce4 MX 440 with AGP8x (NV18) .

NVIDIA GeForce4 Ti 4200 with AGP8x (NV28)

NVIDIA GeForce4 MX 440 with AGP8x (NV18)

NVIDIA GeForce4 Ti 4200 with AGP8x (NV28)
Samsung memory modules, 4ns access time. The highest frequency is 250 (500) MHz, but by default the memory runs at 256 (513) MHz. The chip is clocked at 250 MHz
NVIDIA GeForce4 MX 440 with AGP8x (NV18)
Samsung (BGA) memory modules, 3.6ns access time. The highest frequency is 275 (550) MHz, but the memory runs at 256 (513) MHz. The chip is clocked at 275 MHz

While the NV28 (in Ti 4200) still comes with ordinary memory modules, the NV18 in MX 440 is a different case. Memory chips in the BGA package could be seen only on the MX 460 cards, and the new MX 440 with AGP8x will have them as well. These chips have 3.6ns access time which is usually required for the memory running at 275 MHz, but here the memory is clocked at 256 MHz (a sweet pie for overclockers :-).

NVIDIA GeForce4 Ti 4200 with AGP8x (NV28)

NVIDIA GeForce4 Ti 4200 (NV25)

The design of the NV28 is identical to the reference one for the Ti 4200. The only difference is a cooler mounted on the chip:

Such cooler is often installed on the reference cards of the GeForce3 line. Remove it and you will see the GPU:

The photo above clearly shows that the AGP 8x is supported.

Now look attentively at the NV18.

NVIDIA GeForce4 MX 440 with AGP8x (NV18)

NVIDIA GeForce4 MX 440 (NV17)

Here there are more differences: the memory comes in the BGA package, all memory chips are on the front side and the card supports two heads, i.e. nView, so that an image can be seen on two monitors and/or TV. Besides, the new MX 440 works at a higher frequencies than its predecessor (memory clock speed).

Although the chip's frequency is 5MHz higher (275 against 270 MHz), NVIDIA recommends using a cooler without a fan:

If you remove the heatsink you will see that the chip is marked as 8x:

Test system and drivers

Testbeds:

Pentium 4 based computer (Socket 478):

Intel Pentium 4 2530;
ASUS P4S8X (SIS 648);
512 MB DDR SDRAM PC3200;
Seagate Barracuda IV 40GB;
Windows XP Professional Service Pack 1.

The test system was coupled with ViewSonic P810 (21") and ViewSonic P817 (21") monitors.

In the tests we used the NVIDIA's drivers 40.41 (with a separately received altered INF file).

The following cards are taken for comparison:

Albatron GeForce4 Ti 4200 (250/256 (513) MHz, 128 MB, driver 40.41);
Leadtek WinFast A170T (GeForce4 MX 440, 270/200 (400) MHz, 64 MB, driver 40.41);
Hercules 3D Prophet 9000 Pro (RADEON 9000 Pro, 275/275 (550) MHz, 64 MB, driver 6.166).

Driver settings

In the control panel of the drivers we switched off VSync for Direct3D and OpenGL, as well as AA. In the games' settings we disabled texture compression (when possible). On the SIS 648 based system we used AGP drivers 1.11. Both NV28 and NV18 cards worked in the AGP 8x mode, which was also demonstrated by the PowerStrip and RivaTuner. All the other cards used AGP4x.

Test results

2D graphics

Coupled with the ViewSonic P817 monitor and BNC Bargo cable the tested samples performed brilliantly at the following resolutions and frequencies:

NVIDIA GeForce4 Ti 4200 with AGP8x (NV28)	1600x1200x85Hz, 1280x1024x100Hz, 1024x768x120Hz
NVIDIA GeForce4 MX 440 with AGP8x (NV18)	1600x1200x75Hz, 1280x1024x100Hz, 1024x768x120Hz

3D graphics, 3DMark2001 - game tests

3DMark2001, 3DMARKS

Although we tested the cards on the GeForce4 Ti 4200 with AGP8x and Ti 4200 under different loads (with AA and anisotropic filtering), they got identical results. It means that for 128MB cards the increased throughput of AGP doesn't make much sense. It also concerns the NV18. You will know why a bit later when you see volumes of textures used on one or another 3DMark2001 test.

The new MX 440 which comes with a faster memory gets a decent performance boost, though the RADEON 9000 Pro is still too far ahead thanks to its mechanisms of operation with shaders (it means that the Game4 brings in good scores).

3DMark2001, Game1 Low details

Test characteristics:

Rendered triangles per frame (min/avg/max): 19773/33753/143422
Rendered textures per frame with 16 bit textures (min/avg/max): 7.5/8.8/16.5 MB
Rendered textures per frame with 32 bit textures (min/avg/max): 15.1/17.7/30.3 MB
Rendered textures per frame with texture compression (min/avg/max): 10.7/12.2/21.0 MB

Neither for the NV28 nor for the NV18 the AGP8x renders any help. Look at the texture volume - 30 MB maximum, that is why 64MB on the card is enough to put all the textures into the local memory.

3DMark2001, Game2 Low details

Test characteristics:

Rendered triangles per frame (min/avg/max): 46159/51440/147828
Rendered textures per frame with 16 bit textures (min/avg/max): 8.0/8.8/10.1 MB
Rendered textures per frame with 32 bit textures (min/avg/max): 15.6/17.2/19.8 MB
Rendered textures per frame with texture compression (min/avg/max): 9.3/10.9/13.5 MB

Well, the only thing that should be noted is that the RADEON 9000 Pro outscores even the GeForce4 MX 440 with AGP8x which runs at the higher frequency.

3DMark2001, Game3 Low details

Test characteristics:

Rendered triangles per frame (min/avg/max): 16681/21746/39890
Rendered textures per frame with 16 bit textures (min/avg/max): 2.8/4.1/4.7 MB
Rendered textures per frame with 32 bit textures (min/avg/max): 5.7/8.2/9.4 MB
Rendered textures per frame with texture compression (min/avg/max): 5.0/7.2/8.4 MB

Well, the AGP's 8x mode is not needed at all because of such amount of textures.

3DMark2001, Game4

Test characteristics:

Rendered triangles per frame (min/avg/max): 55601/81714/180938
Rendered textures per frame with 16 bit textures (min/avg/max): 14.9/17.4/20.7 MB
Rendered textures per frame with 32 bit textures (min/avg/max): 28.4/33.5/40.0 MB
Rendered textures per frame with texture compression (min/avg/max): 28.4/33.5/40.0 MB

This test could give a chance to the NV18 card to show capabilities of the higher throughput of the AGP 8x, but unfortunately, this card lacks for pixel shaders (in contrast to the RADEON 9000 which has a similar price). As for the NV28, its 128 MB is a big hanger which can easily house 40 MB of textures.

So, if the new GeForce4 MX 440 with AGP8x sell at the prices the MX 440 is available today, they will definitely be popular with users, because their speed has grown considerably but at the expense of the faster memory, not thanks to the AGP 8x.

The 3DMark2001, like all modern games, mostly deals with video card's memory, not with pumping data through the AGP, and as we found out the video memory size is quite enough for such tests.

3D graphics, game tests

For the performance estimation we used:

Return to Castle Wolfenstein (MultiPlayer) (id Software/Activision) - OpenGL, multitexturing, Checkpoint-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, Grand Cathedral demo, test settings: quality, S3TC OFF

Quake3 Arena v.1.17 (id Software/Activision) - OpenGL, multitexturing, Quaver, 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

Comanche4 Benchmark Demo (NovaLogic) - Direct3D, Shaders, Hardware T&L, Dot3, cube texturing, highest quality

Unreal Tournament 2003 Demo (+Patch 1080) (Digital Extreme/Epic Games) - Direct3D, Vertex Shaders, Hardware T&L, Dot3, cube texturing, default quality, FlyBy CITADEL and FlyBy ANTALUS tests

Code Creatures Benchmark Pro (CodeCult) test demonstrates operation of cards in the DirectX 8.1, Shaders, HW T&L.

AquaMark (Massive Development) test demonstrates operation of cards in the DirectX 8.1, Shaders, HW T&L.

RightMark Video Analyzer v.0.4 (Philip Gerasimov) - DirectX 8.1, Dot3, cube texturing, shadow buffers, vertex and pixel shaders (1.1, 1.4).

Quake3 Arena, Quaver

The GeForce4 MX 440 with AGP8x aces everyone in this test (RADEON 9000 Pro is hopelessly behind).

Serious Sam: The Second Encounter, Grand Cathedral

The screenshots of settings of this game can be found in this review.

At last, the AGP is enabled to a much greater extent for the MX 440, and its 8x mode is now able to bring higher scores. But in case of the NV28 the 128MB memory remains very huge even for this game. Moreover, the RADEON 9000 Pro easily outdrives both the NV18 and MX 440. The programmers from ATI did their best to improve operation of the drivers, in particular, for this benchmark.

Return to Castle Wolfenstein (Multiplayer), Checkpoint

The games under the OpenGL (except Quake3) use the AGP to a more degree, and in this case we can see that the NV18 wins the battle with the MX 440, especially in 1600x1200. Like in the Quake3, the RADEON 9000 Pro loses to its competitors, though here it looks better than in that benchmark.

Code Creatures

No comments.

Comanche4 DEMO

It's interesting that this very processor-dependent test doesn't neglect higher bandwidth of the AGP, and the NV18 boost is noticeable. As for the NV28, I don't see any difference between the NV28 and NV25.

The RADEON 9000 Pro takes the palm again; it falls behind the NV18 only in 1600x1200.

Unreal Tournament 2003 DEMO

CITADEL

The NV18 unexpectedly falls behind the MX 440, probably because of the drivers. Anyway, the RADEON 9000 Pro remains a champion among mid-level cards. As for the NV28, there is no difference between the AGP4x and AGP8x cards.

ANTALUS

The situation is similar but for the fact that the MX 440 is able to beat the RADEON 9000 Pro.

AquaMark

Well, this is the only case when the NV28 outdoes the Ti 4200 in 1024x768; the situation is in general the same.

Although the NV18 managed to outrun the MX 440, the RADEON 9000 Pro comes with more perfect scores (shaders...) .

RightMark 3D

Although this test is very difficult (it was developed to estimate operability of modern accelerators with DX8.1 enabled and the speed of operation), 128MB of the GeForce4 Ti 4200 with AGP8x is more than enough, that is why we notice no difference from the Ti 4200.

Like in the 3DMark2001, for the 128MB cards the AGP8x support means nothing. There are no games which would benefit from a higher bandwidth of the AGP.

The NV18 has received some advantages, but they are not weighty enough. A stronger argument for the MX 440 is faster memory which provides a performance gain as compared with the former MX 440. But anyway, the ATI's RADEON 9000 Pro is still leading.

Conclusion

So,

The less the memory of the video card, the less the effect of the AGP8x. You shouldn't expect any performance increase from the 128MB Ti 4200-8x. But you should account for the fact that NVIDIA has released the Ti 4200 with AGP8x bundled with 4ns memory working at 256 (513) MHz. As you remember, all 128MB cards based on the Ti 4200 were initially clocked at 222 (444) MHz (memory). The Californian company urges the manufacturers to increase memory frequencies on 128MB cards. We didn't compare Ti 4200-8x (256 MHz memory) with old Ti 4200 with the memory working at 222 MHz because it's not correct. Anyway, we have carried out an objective comparison of two cards running at equal frequencies.
In case of the NV18, we compared the MX 440-8x on former 270/200 MHz with the MX 440, as well as revealed performance of the MX 440 with AGP8x at its rated frequencies of 275/256 MHz. Given to the fact that NVIDIA has a strong rival in this market sector, it was very interesting for us to find out whether the new NV18 could fight against the ATI RADEON 9000 Pro. Well, the NV18 yields to the latter quite often. Moreover, it's deprived of pixel shaders, EMBM, and a higher-level anisotropic filtering.

If the new MX 440 with AGP8x are going to sell at the prices of former MX 440, the NV18 stands a chance to become popular. Otherwise, it would be a real problem to fight against the RADEON 9000 Pro which is gradually falling down in price.

The NV28 must certainly be comparable in price with the current Ti 4200. I think that advantages of the AGP8x won't bring tangible benefit as compared with the AGP4x.

BUT! If, even next year, games start using texture volumes over 100-150 MB for a scene or the increased AGP throughput is used more intensively, the current NV18/NV28 will get a second wind and be able to lift up their performance level (by that time the Ti 4200 will probably become a low-end solution).

Andrey Vorobyev (anvakams@ixbt.com)
Alexander Medvedev (unclesam@ixbt.com)

Write a comment below. No registration needed!