Introduction

At the beginning, everything was easy in SLI. Who is wanted to build a dual-GPU system, needed two identical graphics nVidia cards as well as a motherboard with the nForce 4-chipset by nVidia and could already start the fun. When the competing technology from ATi, namely CrossFire, the market reached, the whole situation was already more complicated. At ATi, it not enough to buy two equivalent ATi graphics cards and a CrossFire suitable motherboard with the RD480 chipset. Instead, it was necessary to have a so-called master card in addition to a conventional ATi GPU. This could mix the image of the second video card thanks to the “compositing engine” with which the first to send then full frames at the monitor output.

These boards had however – at least marketing technology – a great disadvantage. So not enough PCIe lanes were available to both chipsets to the two physical PCIe-anzusprechend x 16 ports for the GPUs with logical 16 PCIe lanes per port. The video cards were operated in both cases with eight lanes, which halved the bandwidth accordingly. Some time later nVidia brought the nForce4 SLI x 16 on the market, which was identical to the previous chip, but – you can hear and amazed – by an additional chip brings additional PCIe lines 18 (AMD) or 20 (Intel), which both graphics cards slots now in SLI mode with 16 lanes can be controlled.

ATi had of course no slouch himself, and produced the RD580 chipset, mostly Xpress3200 diligently called, which has not only two logical full PCIe slots, but is also the possibility the GPUs, directly with each other without to communicate directly over the CPU. But one wonders in principle whether doubling the PCIe lanes at all of use is, after all was already at that time hardly a difference between AGP 4 x and AGP 8 x exist.

Now you admit themselves as honest buyers, however, that the demands on the mainboard of SLI and CrossFire configurations, and in particular by the SLI-AA and SuperAA modes are greatly increased. 1600 Communicates in addition only via the PCIe bus in CrossFire mode – other configurations refer Radeon X XT from ATi the compositing engine or the SLI bridge, to relieve the PCIe bus. Therefore the question must be clarified new, whether it is now makes sense to buy the new chipsets with 32 lanes of PCIe graphics cards or if an older chipset with 16 lanes is sufficient. Armed with the four commonly used chipsets, as well as two GeForce 7800-GTX-512- and X 1900 cards and two “Radeon X 1600 XT” we will go projects this question to the bottom.

Technical overview

Old acquaintances had to serve as test candidates for this article: on pages by nVidia the Gigabyte GA-K8NXP-SLI with the nForce 4-SLI chipset of the challenge faced, while the ASUS A8N32-SLI Deluxe as a bigger brother with the nForce 4 SLI x 16 is taken. He was on the side of the ATi-and CrossFire-enabled chipsets the Sapphire PC-A9RD480 CrossFire with the RD480 chip was a duel while the ASUS 8R32-MVP Deluxe as a RD580 colleague attacked the large with the nForce Chipset ASUS Board.

Now we go with the four mentioned boards on the PCIe bus closer to the detail. The GigaByte GA-K8NXP-SLI is an exponent of traditional nForce 4-chip, which provides 20 PCIe lanes total. While 16 are assigned to the two graphics card slots, while two more lanes of PCIe-x 1 slots are used. If a SLI system on the GigaByte mainboard is used, two slots are not 16 logical PCIe lanes to connect the graphics card ready, but only half, say eight lanes per graphics card, because not enough PCIe lanes are available, to operate both slots with the maximum configuration. This halved the bandwidth of PCIe port in the back as well as in the reverse direction.

Similarly, it is taken the direct competitors of RD480 of ATi, which is being used among other things on the Sapphire PC-A9RD480 CrossFire. The chipset can address 22 PCIe lanes, which it is already clear that no full x 16 link to two graphics ports is possible. Both slots are addressed in a CrossFire system with eight lanes. Two additional PCIe lanes used in the Board for the connection between North and South Bridge, while the remaining free can be distributed by the manufacturer of the Board.

The current flagship of the nVidia is the nForce 4 SLI-x 16 chipset, which lingers on the ASUS A8N32-SLI Deluxe. A total whopping 38 PCIe lanes can be distributed to the Board by the manufacturer, even though the motherboard uses mainly the features of old nForce 4-SLI chipset. As a special treat, nVidia used a chip, which the customer additional 18 (AMD) or 20 (Intel) offers PCIe lanes. This additional chip listens to the name “SPP” (System Platform Processor) and is probably connected with the HyperTransport Protocol on the Northbridge. Accordingly both graphics card slots can be addressed logically 16 PCIe lane.

ATi takes a different approach with the RD580 chip. The ASUS 8R32-MVP Deluxe relies on the latest piece of Silicon of the Canadians, which can manage 40 PCIe lanes. This is achieved by an additional chip in but not like in nVidia, but all in a single Northbridge is located. Thus provides 32 lanes for the graphics card slots in a CrossFire system. In addition both graphics cards to communicate on a RD580 boards directly, what especially for CrossFire systems without the compositing engine, such as when two Radeon-X 1600 XT cards a lot of performance to bring.

To maximum on the PCIe bus, we test the individual cards as well as the dual-GPU systems in 1600 x 1200 resolution and a second pass with additional 4 x anti-aliasing and 16-time anisotropic filtering. The dual-GPU system must meet and also the special SLI-AA – respectively SuperAA modes, which represent a burden for the PCIe bus.

Test system

Test system:

• Processor
  • AMD Athlon 64 4000 + (San Diego core, 90 nm, SSE3, 1024 kB level 2 cache)
• Motherboard
  • Gigabyte GA-K8NXP-SLI (nVidia nForce 4 SLI)
  • ASUS A8N32-SLI Deluxe (nVidia nForce 4 SLI x 16)
  • Sapphire pure CrossFire PC-A9RD480 (ATi RD480)
  • ASUS A8R32-MVP Deluxe (ATi RD580)
• Memory
  • 2 x 512 MB crucial BallistiX DDR400 (2-2-2-5)
• Graphics cards
  • ATi Radeon X 1900 XTX (650 / 775)
  • ATi Radeon X 1900 CrossFire Edition (625 / 725)
  • ATi Radeon X 1600 XT (590 / 690)
  • nVidia GeForce 7800 GTX 512 (550 / 850)
• Peripherals
  • AOpen AAP-1648Pro-DVD drive
  • Samsung S-ATA 2 HDD 200 GB disk space (NCQ enabled)
• Driver versions
  • nVidia ForceWare 82.12
  • ATi catalyst 6.3
• Software
  • Microsoft Windows XP Professional SP2
  • Microsoft DirectX 9.0 c

Benchmarks

The following benchmarks were used during our test:

• Synthetic benchmarks:
  • 3DMark05 version 1.2.0
  • 3DMark06 version 1.0.2
  • Aquamark 3
• Game benchmarks:
  • Age of Empires 3
  • Splinter Cell: Chaos theory
  • Fear
  • Serious Sam 2 demo
  • Doom 3
  • The Chronicles of Riddick
  • Call of duty 2
  • Battlefield 2
  • Quake 4
  • Half-life 2: Lost coast

All benchmarks run with maximum details, so that the graphics card will be charged as high. As the setting we decided this for 1600 x 1200. So we pay tribute to the modern high end Accelerator, make the resolutions by their computing power lower than 1280 x 1024 CPU limited.

Attention:After careful consideration and multiple analysis of even recorded game sequences, we have come to the conclusion in the ForceWare driver for nVidia cards increase the quality settings on high quality, because we can effectively combat the texture flickering only with this setting. In addition, this mode is similar to the setting “catalyst a.i. standard” on the ATi counterparts, resulting in largely a tie in image quality.

Driver settings: nVidia video cards

• Performance: High quality
• Vertical sync: off
• Force MipMaps: no
• Trilinear optimization: from
• Anisotropic mip filter optimization: from
• Optimization of anisotropic pattern: from
• Negative LOD bias: clamp
• Gamma adjusted AA (G70): A
• Transparency AA (G70): Off

Driver settings: ATi video cards

• Catalyst a.i.: Standard
• Mipmap detail level: high quality
• Wait for vertical refresh: always off
• Adaptive Anit-aliasing: off
• High quality AF: off
• Truform: Always off

Synthetic benchmarks

3DMark05

• The 3DMark05 is technically still at a very high level. So large textures with the resolution come of shader model 3.0, 2.x or 2.0, use 2048 x 2048, mixed with the use. The last year published program uses complex lighting effects, dynamic shadows, elaborate bump mapping effects and requires above all a high performance of geometry. As a result only the speed of the graphics card reflected, because it is always the bottleneck even in current hardware. The spacious driver controls of all current graphics card manufacturers are the probably the biggest drawback to the 3DMark05. They go so far as that the final results of change depending on the driver in the two-digit percentage range, thus reducing quality optimizations cannot be excluded. Also the synthetic benchmark based which is why he is not a real situation on any game engine. There are more details about this program in our detailed article.

• Download: 3DMark05

3DMark06

• The well-known series of benchmark from Futuremark is now published in version 2006 and consequently listens to the name “3DMark06″. Measure the performance of the graphics card four sequences of the six test scenes and show a graphic splendour, that their same looks. To achieve that the Finns set which is why not only massively the shader model 3.0 is used, extremely elaborate textures, spectacular particle effects, complex shadow computations to cutting-edge 3D technology, and as another highlight “High Dynamic Range Rendering” – short HDRR – are used. While Futuremark opts for FP16 HDR, which currently provides the best possible image quality, but is also expensive to calculate. Thus graphics cards without FP16 blending units, including the X to 0 series from ATi, two test scenes not run 8 x, which is why the number of points of these GPUs generally low fails. In addition, only graphics cards, can the MSAA to a FP16 Rendertarget can run, calculate the HDRR sequences with anti-aliasing. Graphics cards without this ability produce no score in use of anti-aliasing, and are therefore not included. There are more details about this program in our detailed article.

Aquamark 3

• We now come to the last synthetic benchmark in our test course. The program developed by massive development uses an enhanced version of the graphics engine from the U-boot-game AquaNox 2 – revelation. The engine the name running with scissors and supports now also pixel shader 2.0 effects. In addition come still pixel shader of version 1.1 and 1.4, also the Vertexshader 1.1, used. This according to the developers of the first DirectX 9-capable reality benchmark in the world to be enriched with some nice effects such as for example an improved particle system.

• Download: AquaMark3.com

Game benchmarks

Age of Empires 3

• Hardly a strategy game as “Age of Empires” and its successors have made history among the PC players due to their successes. Now, the third part of the series, which appears with a breathtaking strategy title graphic that supports most modern techniques such as the shader model 3.0 and even high-dynamic-range-rendering to appears. We play from a one-minute sequence for the measurements, which corresponds to the real game play and shows several explosions and a fight with many units. All the details are set to the maximum, which also HDRR is used. However, those modes are not equivalent, as for nVidia graphics cards-quality FP16 HDR is used, while the ATi counterparts of the X 1000 series provide only INT10 HDR. Normal MSAA at the nVidia GPUs to work on the ATi cards is accordingly a 1 on the other hand, 5xSSAA mode used. However, the picture quality is similar to that of the normal gameplay. The X 8 x 0-series can in age of Empires 3 no HDRR representing (shader details can not on “Very high exposed”), which is why the performance from the ground from is significantly higher.

Battlefield 2

• Battlefield is most certainly one of the most popular and most played multiplayer games of all time. The successor of Battlefield 2 builds on the success and a spiffy graphic combined with a relatively simple, but very funny game concept. The graphics with relatively modern shader effects, lives but mostly through on few textures as well as a convincing particle and smoke system, producing a dense atmosphere. The details are provided for the measurements to the maximum and we use the tool “BF2Bench”, because only that produces realistic and reproducible results. The several minutes long Timedemo shows several tank, aircraft and military battles from a free camera view.

Call of duty 2

• The Weltkriegsshooter of “Call of Duty 2″ is also the graphics not only with a dense atmosphere and a lot of fun, white to like. So, a completely new graphics engine was written for the game, in which the developers have drawn many “graphics register”. So, sets the game to many shader effects, and thanks to the excellent textures and the very good facial animation is a feast for the eyes. Most impressed the smoke and fog representation that is truly unique in the first-person shooter – so far there such a realistic smoke in any game. But the splendour of the graphic takes its toll on the accelerator and eats the available resources of the GPU like the breakfast. There is also a call of duty 2 of one of the first games that can benefit from a 512 large VRAM. The Timedemo selected by us shows a section from the “Russian campaign”, which is extremely Hardwarefordernd due to the appearance of snow, and the landscape. Several shootout and smoke grenades are so the Timedemo very well for a test course is with of the lot.

Doom 3

• Fear? Shock? Darkness? Graphic flowers? All this is probably there to meet in the Gruselshooter of doom of 3. John Carmack, one of the Chief developer of the game and a legendary figure, when it comes to spectacular graphics engines, has devoted the greatest attention the stencil shadow in his latest work. The entire game is accordingly dark, so that the template-like shadow work well on the player. But there were not enough effects for the developer ID software. So, makes Doom 3 also use by the pixel shader units of graphics cards and also heavily relies on bump mapping, as well as normal maps. Though the textures are improved, but nevertheless is part of Doom 3 to the most demanding titles of 2004 and is thus ideally suited for our benchmark course. The game does ID typically not on DirectX API, as on OpenGL.

Fear

• Doom 3 gets competitive – and what for one! The programmers of new Gruselshooters f.e.a.r.seem to Doom 3 as a great model to have selected, which you seem to do better but almost everything. Among other things, the very oppressive atmosphere by a graphics quality is achieved, that of their peers is looking for. Shader in masses, wonderful bump-mapping, shadows very spectacular, detailed textures, as well as pretty-looking particle effects and much more Gets the player to face why f.e.a.r. already mandatory for a good benchmark course has become. We now use the full version, which has a built-in function of the benchmark for these purposes. That shows a battle and a larger explosion, has been introduced by a freely moving camera. The details are, set, the soft shadows with exception to the maximum.

HL2: Lost coast

• Half-life 2 has been probably no doubt due to his legendary predecessor, one of the most anticipated games of all time. Now it is here and excited not only in playful ways, but also by its graphic, which is made possible by massive “shader model 2.0″-use, among other things. A few months after the appearance of valve brought the free technology demo “Lost coast” on the market, which as a special high-dynamic-range rendering support and offers not only a significantly higher amount of light and lighting dynamics, but also promotes the hardware up to the limit. Valve however has designed for compatibility with older graphics cards and is therefore a “inferior” form of HDRR, that provides not the best picture quality. So are the textures in FP16 format before – or INT16 for video cards without FP filtering – although, however, valve waives FP16 blending. For this reason, also X can represent 8 x 0 cards in lost coast HDRR. The itself created Timedemo shows several firefights with soldiers and a helicopter and impressively illustrates the optical gain by HDRR.

Quake 4

• The known Quake from ID software is each time a highlight for a “first person shooter” fan, because the games offer not only a high entertainment value, but also comes along with a graphic beauty which is often a good deal higher sets the standard. The current version of Quake 4, however, software was programmed by Raven and uses a slightly enhanced Doom 3 engine. Thus, is the graphic at a high level, can set but no new standards. Nevertheless, some for the eye offers play with on a character textures and many shadow plays. The selected Timedemo shows several firefights and spectacular shadow and color games. A bug that would be worth noting the in the Timedemo function of Quake 4 has slipped, which was not yet present in Doom 3. No character shadow appear, as well a great of weapons effects are missing. Accordingly, a piece in our results is the level of frames per second in the correct mode.