Testing NVIDIA SLI (GeForce 6800/6600 PCI-E). Part 2

Introduction, video cards' features
Testbed configurations, benchmarks
Test results: Unreal Tournament 2003
Test results: TRAOD
Test results: FarCry Research
Test results: FarCry Regulator
Test results: FarCry Pier
Test results: Half-Life2, ixbt01
Test results: Half-Life2, ixbt02
Test results: Half-Life2, ixbt03
Test results: DOOM III
Test results: 3DMark05 Game1
Test results: 3DMark05 Game2
Test results: 3DMark05 Game3
Test results: 3DMark05 MARKS
Conclusions

We proceed with our tests of "duets" of video cards based on GeForce 6800/6600.

In this part of the material we'll analyze the work of such tandems based on the new NVIDIA product – nForce4 SLI. ASUSTeK was practically the first manufacturer to release such production motherboards on the market, so we have carried out our tests on the basis of the A8N SLI Deluxe motherboard.

You can get more details about this motherboard from the corresponding article in the Motherboards section of our web site. I will just remind you that nForce4 was the first chipset to offer not only support for modern AMD solutions but also simultaneous use of two PCI-Express x8/16 video cards, which means that such motherboard can have two PCI-E slots.

Of course not all nForce4 chipsets support this feature, that's why the series: nForce4, nForce4 Ultra, nForce4 SLI. And only the last one supports two-slot video card solutions, and consequently SLI.

So, our testbed got a new look, it has become dark-brown (the color of the motherboard PCB) and acquired a fashionable illuminated cooler from Gigabyte.

The main heroes are the same, though – video cards :), so let's get right to them.

Video Cards

Actually, there is little sense in describing them, because we used the same accelerators as in the first part of our material.

Installation and Drivers

Testbed configurations:

Athlon 64 (939Socket) based computer
- AMD Athlon 3500+ CPU (overclocked up to 2400MHz) (L2=512K);
- ASUS A8N SLI Deluxe motherboard based on NVIDIA nForce4 SLI
- 1 GB DDR SDRAM 400MHz
- WD Caviar SE WD1600JD 160GB SATA HDD
Operating system – Windows XP SP2; DirectX 9.0c
Monitor: Mitsubishi Diamond Pro 2070sb (21").
NVIDIA drivers 66.93/ATI - 6.497.

VSync is disabled.

A pair of GeForce 6800GT PCI-E video cards

A pair of GeForce 6600GT video cards

Note that the ASUS product provides for selection of motherboard operating modes: two video cards or one, this can be done with the help of a special module between the PCI-E slots, which resembles SO-DIMM. The module has two sides and works as a switch: insert it one side into the slot – the first mode is enabled, insert it the other side – the second mode will be enabled. If you need to activate two video cards, each one will operate as PCI-E x8. If you switch to a single video card mode, the video card in the first slot will get PCI-E x16.

There is an additional power connector right over the first PCI-E slot: you should plug the cord from the power supply unit into this connector in case of SLI mode (this is a requirement set by designers to increase operating stability).

The first slot is marked blue, a monitor is connected to the video card inserted to this slot. Another nuance: the slots are placed farther from each other than NVIDIA intended, so the SLI connector from this company will not do in this case: there are two slots between these video cards instead of only one. However ASUS designed a similar proprietary SLI connector (synchronizer) and ships it with the motherboard (you can clearly see it on the photos).

Test results: performance comparison

We used the following test applications:

Unreal Tournament 2004 v.3325 (Digital Extreme/Epic Games) – Direct3D, Vertex Shaders, Hardware T&L, Dot3, cube texturing, max quality; demo (MOON)
Tomb Raider: Angel of Darkness v.49 (Core Design/Eldos Software) – DirectX 9.0, Paris5_4 demo. The tests were conducted with the quality set to maximum, only Depth of Fields PS20 was disabled.
Half-Life2 (Valve/Sierra) – DirectX 9.0, demo (ixbt01, ixbt02, ixbt03 The tests were carried out with maximum quality, option -dxlevel 90, presets for video card types are removed from dxsupport.cfg.
FarCry 1.3 (Crytek/UbiSoft), DirectX 9.0, multitexturing, 3 demo from Research, Pier, Regulator levels (the game is started with -DEVMODE option), Very High test settings.
DOOM III (id Software/Activision) – OpenGL, multitexturing, test settings – High Quality (ANIS8x), demo ixbt1 (33MB!). We have a sample batch file to start the game automatically with increased speed and reduced jerking (precaching) d3auto.rar. (DO NOT BE AFRAID of the black screen after the first menu, that's how it should be! It will last 5-10 seconds and then the demo should start)
3DMark05 (FutureMark) – DirectX 9.0, multitexturing, test settings – trilinear,

Unreal Tournament 2004

Diagrams with results

6800Ultra SLI versus 6800Ultra/RADEON X850 XT PE – it's clear that we can get maximum performance gain, which is very good (up to 75%), only with maximum video card load. What concerns the competition with the most powerful accelerator from ATI for these days, there is almost twofold advantage in the heaviest mode with AA and AF. Quite impressive.
6800GT SLI versus 6800GT/RADEON X850 XT PE – almost the same picture, but the performance gain from SLI is higher (it's clear: the weaker a card is, the less it's limited by CPU, the more contrasting the results are); but concerning X850 the advantage is a tad smaller, but still it remained twofold.
6600GT SLI versus 6600GT/RADEON X800 XL/RADEON X800 XT – SLI gains are more moderate, which is rather strange. Perhaps the problem is in drivers. One should say that it may be the reason 6600GT SLI did not sustain competition with X800 XL (which is cheaper than two 6600GT cards).

TR:AoD, Paris5_4 DEMO

Diagrams with results

6800Ultra SLI versus 6800Ultra/RADEON X850 XT PE – everything is peachy! SLI gains, advantage over the competing X850 XT PE. But still two 6800Ultra cards are much more expensive than one X850.
6800GT SLI versus 6800GT/RADEON X850 XT PE – the same picture, only percentage values are lower relative to X850, and the SLI efficiency is even higher than in the previous case.
6600GT SLI versus 6600GT/RADEON X800 XL/RADEON X800 XT – the same! And even X800 XL is left lagging far behind.

Here is a striking example of a test, for which NVIDIA drivers were fine tuned. The lion's share of performance in this test is up to GPU, and practically pure SM2.0, which makes it easy to fine-tune software.

FarCry, Research

Diagrams with results

6800Ultra SLI versus 6800Ultra/RADEON X850 XT PE – the performance gain from SLI is not that large even with AA+AF (performance turns on the system resources). As you can see, the advantage over X850 XT PE is also reduced relative to the previous test
6800GT SLI versus 6800GT/RADEON X850 XT PE – it's quite logical: the card is a tad weaker and so its resources are exhausted faster in heavy modes, and the effect of SLI is correspondingly much greater, though the competition with X850 XT PE is faded.
6600GT SLI versus 6600GT/RADEON X800 XL/RADEON X800 XT – the advantage does not exceed 60%, so we again draw a conclusion that the problem is in drivers. But the sweet couple of 6600GT is bested by X800 XL, the ATI product being faster and cheaper at that.

FarCry, Regulator

Diagrams with results

The results are practically similar to the previous ones. Obviously, the performance turns on the system resources and that does not allow large yield from SLI (but these are the conditions. And potential buyers of such systems shouldn't forget about it: having paid about 1000 USD for a couple of video cards, you will not necessarily get stunning performance gains even in the heaviest modes, because the lion's share of games will turn on the capacities of CPU and FSB long before the video card capacities are depleted).

FarCry, Pier

Diagrams with results

And this test demonstrates what happens if the game depends much on CPU (or a separate game scene) or when everything is up to geometry calculations, which are not duplicated by SLI. So the effect of SLI usage is minimum here. Moreover, there are many percentage minuses and thus defeats.

Half-Life2, ixbt01 demo

Diagrams with results

The picture is repeated again. Yes, the SLI effect is present, but it's weakened by factors, where SLI is powerless. It's geometry and system resources. Even with 2400 MHz Athlon 64 (this is quite an expensive processor). But the picture gets even bleaker for both SLI tandems, when we see that X850 XT PE, which is certainly less expensive, is often more powerful. To say nothing of the pair of 6600GT, which was completely defeated by X800 XL.

Half-Life2, ixbt02 demo

Diagrams with results

I think there is no need in comments, this test provided almost no good news for fans of new superpowerful SLI systems. There are some gains, but on the whole the competitors from the Canadian manufacturer turned out stronger.

Half-Life2, ixbt03 demo

Diagrams with results

The same picture. And the situation is even worse. On the whole it's obviously useless to buy a SLI system for Half-Life2. Probably the same applies to other games based on this engine.

Perhaps NVIDIA will improve its drivers, but still, the game is so CPU-dependent that even the Californian programmers can do nothing about it.

DOOM III

Diagrams with results

That stiffens the spirit of SLI fans! Yes, this test again demonstrated huge performance gains from SLI reaching doubled values. It's DOOM III, you know – favourite child of the American manufacturer, and its drivers are fine tuned. :) And the advantage over ATI competitors certainly reaches incredible values! We'll see what results will be demonstrated by games based on this engine. But the picture looks optimistic for SLI so far.

Quite naturally, this test and DOOM III are the first priorities to fine tune the drivers for. That's why we see brilliant SLI results, which should be ideally the case for each game not turning on CPU and geometry.

Conclusions

So, final conclusions on two large analyses.

The main conclusion: SLI operates well and is even very effective in applications, which do not turn on CPU and FSB capacities, and also where geometry calculations generally do not slow down GPU performance.

There is still much criticism about the quality. The drivers are far from perfect. Especially it concerns 1600x1200 resolution, where the screen sometimes displays a cacophony of colors instead of the sky image, for example.

Detailed analysis of the performance in time graphs from the first part of our review by the example of 3DMark05 Game1/2/3 made it clear that, for example, Game3 uses the frame-by-frame switching between the cards (as in ATI MAXX between the chips) and so instant FPS changes at great amplitude. That's why I think that I have made the wrong conclusion in the first part about this mode being unavailable. This mode is available, but it can be enabled/disabled only by the driver, users cannot do that.

Fourthly: Obviously, the SLI efficiency completely depends on the given game support by the drivers. That is SLI does not operate well absolutely in all games. The process complexity has grown by orders since 1998 and so very much is up to software. Consequently, we are waiting for guys from NVIDIA to fix the existing bugs and to improve SLI efficiency where possible.

And the last thing: it's clear that to buy two 6600GT video cards for this mode is often inexpedient, X800 XL sometimes turning out advantageous. It's up to users' tastes. It is reasonable when your purchase of 6600GT and then of the second card is only an intermediate solution before you buy a couple of 6800GT or 6800Ultra. Motherboards on nForce4 SLI promise to be not that expensive, you don't have to change RAM (the same DDR400), processors are quite affordable, so these are all pros for SLI. One should also note that the new platform was stable, we had no problems or gripes about it, which is a pleasant surprise.

In our 3Digest you can find more detailed comparisons of various video cards.

Theoretical materials and reviews of video cards, which concern functional properties of the GPU ATI RADEON X800 (R420)/X850 (R480)/X800XL (R430)/X700 (RV410) and NVIDIA GeForce 6800 (NV40/45)/6600 (NV43)

We express our thanks to
"ASUSTeK"
for the provided devices for the testbed

Andrey Vorobiev (anvakams@ixbt.com)

December 21, 2004

Write a comment below. No registration needed!

Testing NVIDIA SLI (GeForce 6800/6600 PCI-E).
Part 2 – NVIDIA nForce4 SLI Platform

Contents