![]() |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| « Back - Print | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
Page 1 - Introduction Little more than two years have passed since NVDIA first introduced their SLI technology. Short for Scalable Link Interface, SLI is the first solution to be able to connect two PCI-Express video cards based on the same GPU in a way that they will operate parallel to each other. This type of approach allows, at least on paper, to have an increase in performance of up to double what you would normally have with a single video card. In the first months of SLI’s release, it potential was limited by driver problems and other software errors. However, forwarding two years into the future and with the release of several new driver versions from NVIDIA, SLI technology has really matured as a whole. Stability has been greatly improved and support has been added to an increasing list of games. SLI support has also been made more accessible to the general public with the release of average and entry level video cards that support the technology. Currently, SLI will even be made available on NVIDIA’s most budget level class of GeForce 7 video cards, the 7100 GS, thus making it a solution that is affordable and available to a vast amount of computer users. At this moment, SLI and Crossfire technologies can be considered the first full phase of maturity for a broader category: multiple GPUs. The next phase is for more than two video cards being linked up to each other in an effort to, yet again, increase performance. As NVIDIA was first to reach the market with a dual video card solution, they have yet again been able to beat ATI in announcing a solution that utilizes more than two GPUs. The technology is, none other than, Quad SLI (the name isn’t very original). Quad SLI has four GPUs operating parallel to each other, theoretically doubling the performance in respect to an SLI setup.
The first public demonstration of this technology from NVIDIA happened January 2006 at the Consumers Electronics Show (CES) of Las Vegas. From that moment on, NVIDIA displayed many more glimpses of Quad SLI at various events, however, its performance at that time was not an accurate example of its true potential due to early driver problems. Initially, Quad SLI was going to be available only in systems of partner OEMs, and NVIDIA, in fact, allowed the technology to be used, exclusively at first, in Dell’s XP 600 Renegade. The XPS Renegade was made available for purchase in March 2006 at the price of $10,000. The system came locked and loaded with a pair of custom NVIDIA 7900GTX 512 MB cards. As can be expected, the systems were extremely limited and available to a very small portion of the public because of their price. Although the XPS Renegade looked cool, we personally preferred going for a solution that wouldn’t have required us to sell our car. The 7950 GX2 was released for retail June 5, 2006. The 7950 GX2 replaced its predecessor the 7900 GX2, which suffered from many problems including layout, size, noise, power consumption, and price. With the release of the 7950 GX2, much of these problems have been addressed and changed. The PCB has been taken down to the size of a regular 7900 GTX and now only a single PCIe power connector is required. Although the GeForce 7950 GTX was compatible with Quad SLI on release, users could not take advantage of this feature until the release of the Forceware 91.45 drivers. With the release of the new drivers, this technology has been made officially available to customers who are passionate enough to want to assemble a Quad SLI system alone. Quad SLI, as we will see in the course of this article, is a technology that traces its roots back to SLI. At the moment, driver compatibility and DirectX 9 are the main limiting factors in addition to cost and availability.
In the course of this analysis we will benchmark our Quad SLI setup that utilizes two 7950 GX2 (four GPUs) video cards against a system armed with two GeForce 7900 GTs running in SLI, a system with two ATI Radeon X1900XTX video cards running in Crossfire and one system with an ATI Radeon X1950XTX video card also running in Crossfire. Page 2 - The Multi-GPU Systems The following table displays the technical characteristics of the four main top of the line video cards from ATI and NVIDIA next to their specifications when they are running in Crossfire or SLI mode.
The first NVIDIA solution we will be looking at today is the GeForce 7950 GX2. The 7950 GX2 comes equipped with two GPUs, thus making it have a total of four under SLI, coining the term Quad SLI. The cards each have two G71 GPUs, the same GPUs are also on the GeForce 7900 GTX cards. The 7950 GX2, however, is set at a lower clock. Technically two boards screwed together, two GPUs of the 7950 GX2 are connected to each other using an on-board two way communications system. From there they both connect to the motherboard using only one PCI-E 16x slot. Quad SLI systems are basically systems configured to run with two 7950 GX2 boards, thus adding up to a total of four GPUs. Combined together, Quad SLI systems hold the most graphical power consumers can get in their systems currently. This, however, does not necessarily mean that it will have the highest absolute performance. Moving on to ATI, we see that their solution, Crossfire, has a different approach to having multiple GPUs running parallel to each other. ATI’s solution requires users to have a Mastercard and an ordinary slave card in order to enable Crossfire. Unfortunately at the moment there is no Mastercard for the X1900XTX, and so we had to revert back to using an X1900XT Mastercard paired with a (slave) X1900XTX. Since the Crossfire architecture always require the Core/Memory frequencies of the two video cards involved to be equal to each other, our X1900XTX slave video card was forced to scale its core and memory clock to that of the X1900XT (this happens automatically). As a direct consequence of this we cannot use the full potential of the X1900XTX and must settle for 625 MHz Core and 1,450 MHz memory.
Our two GeForce systems both use bridges, which are mechanical connectors, for the transmission of data between the multiple GPUs. Every video card in its turn is connected to a PCI-Express 16x slot on the motherboard.
As can be seen, our Crossfire video cards connect externally by using a special cable that comes with the Mastercard. Page 3 - A Brief Guideline to High Resolutions Before moving further into this article, we would like to give a quick and painless introduction to the various resolutions that are available for use with these video cards. In our opinion, what would be the use of Multi-GPU technologies if we didn’t push them to the max by using the highest resolutions possible? As is pretty evident from how high-end they are, and also from their price, multi-GPU systems naturally allow, in addition to AA and AF being enabled, the use of very high resolutions. The diagram above compares various resolution levels. As can be seen, in relationship to the standard 4:3 ratio, the widescreen 16:10 ratio can really create a much more engaging experience inside a typical action game. Because of the larger horizontal length, widescreen aspect ratios can provide a much more complete view in 3D games.
To show a little bit more comparison between widescreen and standard 4:3, the two screenshots show just how much more engaging a 3D game becomes when using wider resolutions. For more information on gaming with wide screens resolutions, we suggest you to visit Widescreengamingforum, an excellent source of information in this matter. The total number of pixels that a video card must process depends on the resolution used. The larger the resolution is the larger the complexity and size of the operations to which the video card will be subjected to:
From the above, we hope that you can more easily understand why managing a resolution such of 2560x1600 pixels, at the present moment the highest resolution the largest 30inch TFT screens available on the market can handle, would be such a daunting task for any video card. The 2560x1600 resolution is pretty much double, in terms of Mpixels, the resolution of 1920x1080 which is the necessary resolution to display HD video. Page 4 - The Quad SLI System In order to construct a Quad SLI test system that any of our readers could build, we chose our components carefully and only bought parts that are available for retail. Our motherboard, the Asus M2N32-SLI Deluxe, is a Socket AM2 based motherboard that comes with the NVIDIA nForce 590 SLI chipset. We used exactly 2 Gigabytes of DDR2-800 memory along with two Point of View 7950 GX2 video cards. The two Point of View cards are identical to the NVIDIA reference card along with all other GeForce 7950 GX2 solutions available on the market. The product conforms exactly to NVIDIA’s recommended settings and comes with a 500 MHz clock for the GPU and a 1.2 GHz clock for the video memory. Each 7950 GX2 has 1024MB Mbytes (512 Mbytes for each core). It is no mystery that all GeForce 7950 GX2 cards are constructed in the same factory, and are supplied from NVIDIA to several partners who sell them as finished products. After installing the video cards, it is noticeable that they are relatively close to each other. The direct consequence is that during operation the video cards generate a remarkable amount of heat, so we suggest making sure there is an adequate cooling system installed in your system in order to avoid overheating problems. The video cards come with a copy of Prince of Persia: The Two Thrones on DVD, a CD that contains drivers, a DVI to VGA adaptor, an external box for management of video signals with connections for component, composite video and S-Video, a Molex to 6 pin PCIe connector, and a video card manual. NVIDIA requires the use of an additional power source for their 7950 GX2 cards in order to guarantee operating stability while connected in Quad SLI. In the course of the tests we used a 620 Watt Enermax PSU that is part of their Liberty series. Using Page 5 - Quad SLI: How it Works The GeForce 7950 GX2 is mounted with two G71 graphics processors. Each GPU comes equipped with 8 vertex shader units (VSU), 24 pixel shader units (PSU), 24 texture mapping units (TMU), and a 256 bit memory bus. Therefore, combined these two GPUs have 16 VSU, 48 PSU, 48 TMU and a 512 bit memory bus.
The distribution of the work load between the two GPUs follows in the footsteps of how SLI technology does so. To be honest, there isn’t much use to buy a GeForce 7950 GX2 unless you intend to buy another one and use it for Quad SLI. If you only have a single GeForce 7950 GX2, the card does not do much other than implement SLI technology on a single video card, which until now has only been used to make two GPUs on two different video cards work simultaneously with each other. The real innovation introduced by the GeForce 7950 GX2 is that it is the first mass produced video card to use a PCI-E switch. The PCI-Express switch between the two PCBs interconnects the two video cards and allows them to communicate with each other using the switch’s 48-lanes. The switch connects 16 PCI-Express lanes to each GPU and the remaining 16 lanes are used to interface with the motherboard’s PCI-Express connector. A Quad SLI configuration is obtained through the use of two GeForce 7950 GX2 cards, totaling 4 GPUs operating simultaneously with each other.
So what are the modes with which NVIDIA operates SLI technology? Two are techniques which were adopted from the first implementation of SLI from the time of the GeForce 6800 cards:
Page 6 - AA and AF through the Driver As we emphasized earlier in the article, the significant power that a Quad SLI has allows for higher resolutions and also offers higher image quality. NVIDIA has introduced three new modes of AA called SLI Anti-aliasing: 8x SLI AA, 16x SLI AA and 32xS SLI AA. These modes improve the total image quality of rendered scenes, and are for this reason, where possible in our testing, turned on. With the introduction of ForceWare 90 Series, NVIDIA’s new control panel allows users to manually set their preferred level of anisotropic filtering and Anti-Aliasing. Even if it is not supported by the specific game, you can force the settings on the game.
Page 7 - The Test Configuration In order for us to build a system to test SLI, Quad SLI and Crossfire technologies, it was necessary for us to build two distinguished test systems. SLI demands use of platforms based on NVIDIA chipsets, while Crossfire can be used with solutions based on ATI or Intel 975X chipsets. We opted on an nForce 590 SLI motherboard for the SLI systems and an ATI Crossfire Xpress 3200 motherboard for the ATI Crossfire video cards. We combined the motherboards with an AMD Athlon 64 FX62 and 2 Gigabytes of DDR2-800 system memory. For more details:
All tests were done starting at a resolution of 1280x1024 pixels and tested up to 2048x1536 pixels. Unfortunately, it was not possible for us to use a resolution of 2560x1600 pixels for the lack of a display that would be able to support the resolution. At the present moment there are very few displays that support 2560x1600. Because these solutions are particularly expensive, availability on the market of these displays is rather thin. The games we will be testing with are:
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||