by Alexey Stepin , Yaroslav Lyssenko, Anton Shilov
02/25/2005 | 02:15 PM
Wise people say there is nothing new under the sun – brilliant ideas are all right here in the air waiting for someone to take and make them a reality. That’s true with computer hardware, too. In an attempt to validate its claim to the title of the manufacturer of fastest consumer graphics cards in the industry, NVIDIA Corporation went what was seemingly the most obvious way. The company suggested that the user united two graphics cards based on specific GeForce6 GPU models into an array that would ensure a much higher performance than that provided by single-chip solutions from ATI Technologies, NVIDIA’s archrival.
<%BANNER[article]%>This “simple” idea gave the user an incentive to transition to PCI Express platforms that would support two graphics cards, but it was also associated with certain requirements like a special mainboard and a high-wattage power supply. As it turned out later, the support of particular games in NVIDIA’s driver was also necessary to get the desired speed boost and to avoid problems like hang-ups and visual artifacts.
Pursuing its promotion plan for multi-GPU solutions, NVIDIA surely expects a growth of its profits, an increase of its market share, and a universal recognition among the users, but the company will also have to optimize the driver for graphics card arrays and to do other word like collaborating with game developers. Anyway, despite all the difficulties, we can say that NVIDIA’s multi-GPU technology called Scalable Link Interface (SLI) has a real chance to conquer the market: the necessary graphics cards and mainboards are already available and can be purchased freely, although at a rather high price.
Today we will study the operation of multi-GPU arrays based on NVIDIA’s GeForce6 graphics cards and the difficulties or possible problems you may encounter with them. We will also offer you the results of our testing the arrays in a wide range of popular computer games.
The history of multi-processor computer graphics goes back more than a decade ago. In 1993 SGI announced the Onyx series of graphics servers of awesome (for that time) performance. The graphics subsystem of those machines consisted of a geometry-processing unit, a display output generator, and the so-called raster manager that processed textures. The cheapest version of such a system, called VTX, isn’t very interesting, although its geometry-processing unit board carried six special-purpose Intel i860XP processors. The more remarkable Reality Engine 2 system permitted to boost the performance by increasing the number of the raster manager units from one to four. The amount of the texture memory didn’t multiply, since its contents were copied for each board, but the frame buffer size in the maximum configuration was as large as 160 megabytes – that was incredibly big at that time.
Thus, the SGI Onyx can be regarded as the first commercial system equipped with a multi-processor graphics subsystem. Such machines cost up to a million dollars, so only wealthy companies could afford one. There was no talking about using such a system for gaming, of course.
Two years later the era of hardware 3D graphics began on the ordinary desktop PC. In 1995, 3dfx Interactive, later devoured by NVIDIA, introduced its first gaming accelerator called 3dfx Voodoo Graphics. That was a revolutionary innovation that improved both performance and quality of computer graphics.
The Voodoo Graphics chipset consisted of two chips: the Frame Buffer Interface (FBI) was responsible for the frame buffer and the Texture Mapping Unit (TMU) processed textures. The chipset could scale the performance up by adding more TMUs – up to three TMUs per one FBI. We don’t know if such configurations ever existed, though. A majority of manufactured Voodoo Graphics cards carried one FBI, one TMU and 4 megabytes of graphics memory (2 megabytes reserved for the frame buffer, and the rest for the textures), but Obsidian3D, a manufacturer of high-performance graphics systems, released the Obsidian 100SB graphics card with two Voodoo Graphics chipsets that were working together in the SLI mode. What did it mean? The abbreviation SLI was fleshed out as Scan Line Interleave. The name described the operational principle of the technology: one Voodoo Graphics chipset was responsible for even-numbered lines of the frame, and the other for odd-numbered ones. Thus, the load was equally divided between the two graphics accelerators and the overall performance would grow up.
The Obsidian 100SB was not a mass product, and the SLI mode for the Voodoo Graphics remained an exotic feature then, but in early 1998 3dfx introduced its next chipset. The Voodoo2 had a higher performance and it became the first mass product to officially support the option of increasing the performance by uniting two Voodoo2-based graphics cards with SLI technology.
Each Voodoo2 graphics card had a connector for attaching to another such card through a special flexible cable. Frankly speaking, this configuration didn’t really take off due to a most trivial reason – the cost of the system was too high. In 1998 one Voodoo2 graphics card would cost more than $300! True enthusiasts did buy two Voodoo2 cards since this duo was unmatched then in terms of performance and graphics quality. There existed a single-PCB alternative from Quantum3D: the Obsidian2 X-24 was even faster than a Voodoo2 SLI configuration, but cost too much, about $650. Quite naturally the latter solution didn’t become widely popular, either.
The progress in the graphics market going on, in the spring of 1999 NVIDIA released its new graphics chip TNT2 whose Ultra version could challenge the speed of the Voodoo2 SLI and could also work with 32-bit color – products from 3dfx lacked the latter feature. Riva TNT2 Ultra graphics cards, although expensive by themselves, would be cheaper than a pair of Voodoo2, without requiring a separate card for the 2D mode.
ATI Technologies also tried to couple two chips on one PCB for a higher performance as it had no other answer to NVIDIA’s release of the GeForce 256 in the fall of 1999. ATI took a different approach than 3dfx Interactive. The point of MAXX – that was the name ATI had given to its multi-GPU technology – was to put two chips on one PCB and make them render different frames simultaneously and then output the frames on the screen alternately. Another name for this technology is Alternate Frame Rendering (AFR).
The performance of the new Rage Fury MAXX graphics card roughly matched the GeForce 256 SDR and was even higher in 32-bit color, but serious defects in the implementation of AFR technology ruined it. Particularly, the two RAGE 128 PRO graphics chips were synchronized by the driver and this negatively affected the latencies when switching between the chips. Quite often the chips fell out of sync and the performance of the card would slump down suddenly or the image would become jerking. Besides that, the peculiarities of MAXX technology prohibited using the Rage Fury MAXX with operating systems other than Windows 98. So, this attempt to issue a multi-processor solution into the desktop graphics market ended in a failure, too.
Meanwhile, 3dfx (the company had changed its name) was all beset with troubles. Its solutions that had been the height of technical perfection just recently were now their competitors’ inferiors from the architectural standpoint. Their performance didn’t seem high anymore, either. For some reason the company again put its stake on multi-processor configurations, although it had already become clear that such graphics cards would be too complex and expensive to be a bestseller.
So, November 15, 1999, 3dfx announced their new VSA-100 chip that supported 32-bit color and could work in multi-processor configurations. This latter feature was dubbed Voodoo Scalable Architecture (VSA). The new incarnation of 3dfx’s SLI had been greatly improved. Particularly, it permitted that up to thirty-two VSA-100 chips be united in a single graphics array, each chip processing a certain sequence of the frame lines. The length of the sequence could change dynamically depending on the load. The “master” chip then formed the final frame out of the line sequences prepared by the “slave” processors.
Alas, actual graphics cards on the VSA-100 chip took too long to come to market – NVIDIA with its GeForce 256 and GeForce2 GTS had firmly established itself there. NVIDIA’s chips were not at all slower but were more advanced from the technical point of view as they featured a hardware T&L unit. So, only the dual-processor implementation of VSA saw the light of day as Voodoo5 5000 and Voodoo5 5500 graphics cards. The four-processor Voodoo5 6000 that required a special external power adapter was manufactured in limited quantities and such cards are real curios nowadays.
There were no other attempts from 3dfx Interactive although it had new products with multi-chip architectures on its roadmap. These plans never came true as at the end of 2000 3dfx was devoured by NVIDIA Corporation. The computer world forgot about consumer-class multi-processor graphics solutions for a while.
September 16, 2003, another attempt to promote multi-processor graphics cards into market was made. XGI, earlier a graphics division of SiS, announced the Volari series of graphics processors that supported DirectX 9.0 and could work in dual-chip configurations. It looked well on paper: two GPUs with eight rendering pipelines in each could perform wonders, but the actual implementation ruined the concept. Having little experience at building such complex systems, XGI created something in-between 3dfx’s VSA and ATI’s MAXX technologies. XGI took the master-slave concept from the former and the alternate frame rendering from the latter. But the two GPUs were connected via the BitFluent bus with a peak bandwidth of only 2.13GB/s. This automatically created a bottleneck in a system of two Volari V5/V8 GPUs.
Besides that, the chips themselves were so weak that even the topmost model, the Volari Duo V8 Ultra (for details please see our review called Club3D Volari Duo V8 Ultra Review: XGI Volari Family Coming to Graphics Market), could only deliver the performance of the RADEON 9600 XT level, but with a horrible quality of the image and at a price of about $500! No wonder the attempt to come to market with such hardware was a complete failure, and the Volari Duo V5/V8 soon sunk into oblivion.
Looking back at the history of consumer multi-chip graphics cards we see only failures and defeats, as every attempt to promote such a product in the market was ultimately unsuccessful. The idea of using several graphics processors in an ordinary desktop PC had seemed to be buried and abandoned forever, but quite recently NVIDIA dug it out and implemented on a whole new level.
First of all, you should know that NVIDIA’s SLI means Scalable Link Interface rather than 3dfx’s “Scan Line Interleave”, so there are fundamental differences between the two technologies, despite the identical abbreviation. On the other hand, we can see some traces of the Voodoo Scalable Architecture in NVIDIA’s SLI. For example, one of the cards in an NVIDIA SLI configuration is a master and another is a slave, and the master is responsible for assembling the final frame and outputting it onto the monitor. Yet there are some significant differences: NVIDIA’s multi-GPU technology uses its own methods of distributing the load between the graphics cards, and these methods differ greatly from 3dfx’s line interleave.
NVIDIA SLI offers two operational modes for a higher performance: Split Frame Rendering (SFR) and Alternate Frame Rendering (AFR). A third mode when only one graphics card is working is added for compatibility with older games.
Split Frame Rendering (SFR) means that the frame is split in two and the two graphics cards each renders its own part of the frame. The gist of the technology is that the frame is not necessarily split in two equal parts – their size may vary dynamically depending on the complexity of the scene. In the latter case the efficiency and performance would be the highest as each graphics card bears almost the same load. NVIDIA refers to this mode as Symmetric Multi-Rendering with Dynamic Load Balancing (SMR). So far they are only talking about uniting two NVIDIA GPUs, but four and more GPUs may be supported in the future.
Alternate Frame Rendering (AFR) means the frames are being rendered alternately by each of the graphics card, similar to ATI’s MAXX.
The SMR mode is, however, the most thrilling thing. According to NVIDIA, the intelligent algorithms of dynamic load distribution make it possible to achieve a performance gain of 80-90 percent in the SLI mode. That’s a brave claim indeed considering that the efficiency of Video Array technology developed by Alienware is from 40 to 70 percent. The downside is that a scrupulous optimization of the drivers and applications is required to achieve such a big performance growth in the SLI mode. We can compare this to computers with multiple CPUs: their performance gain depends on how well the particular application can distribute its threads among different processors. Aside from the topic, we want to note that the number of programs optimized will depend on the popularity of multi-core central processors.
But let’s get back to NVIDIA’s SLI. We will see later on if the claim about an almost double performance gain has anything substantial in itself, but now we will discuss the technical aspects of the implementation of NVIDIA’s multi-GPU technology. Right now the SLI mode is supported by NV45 (GeForce 6800 Ultra/GT) and NV43 (GeForce 6600 GT) graphics processors. The die of each of these GPUs includes a special MIO unit responsible for connection and synchronization of the GPUs.
Considering the fantastically high bandwidth of the PCI Express interface they might have used it to transfer the necessary data instead, but NVIDIA wanted to get the utmost from SLI technology and minimize possible latencies and performance hits by equipping its new GPUs with a dedicated interface. The MIO slot located on the graphics card looks like a PCI Express x1 slot, but somewhat narrower. This slot is probably also responsible for the Frame Lock function, i.e. it avoids the frame tearing effect as a result of GPUs falling out of sync. By the way, the inexpensive mainstream GeForce 6600, although is based on the NV43 core, is officially devoid of SLI: the necessary slot is just not wired on the PCB.
When the graphics cards are plugged into the mainboard’s slots of the PCI Express x16 design, they are then connected with a special adapter. The adapter is a small PCB with two corresponding connectors.
Here’s how it looks when the cards are in the system and joined with the adapter:
It takes meeting a few requirements even to enable the SLI mode, not to mention to make it work correctly. You’ll learn about these requirements in the next section of this review.
For the SLI mode to work some other conditions must be met besides the mainboard’s having two PCI Express slots designed as x16. First, either “16 lanes + 8 lanes” or “8 lanes + 8 lanes” configuration must be supported. Second, the mainboard’s chipset must be supported by the ForceWare driver. So far only two chipsets meet this requirement; they are Intel’s i7525 “Tumwater” and NVIDIA’s nForce4 SLI. It is NVIDIA’s chipset that’s the foundation for building SLI-compatible systems as the i7525 is intended for high-performance workstations with the Intel Xeon processor. We don’t think a dual-processor mainboard priced about $350-500 is likely to appeal to an ordinary user, so we will only discuss SLI technology with regard to mainboards on NVIDIA’s chipsets. Theoretically two graphics cards from NVIDIA can also be united into a SLI duo configured as “16 PCI Express lanes + 4 lanes”, but a certain performance hit may be observed due to the low bandwidth of the second slot.
This didn’t prevent some mainboard manufacturers from releasing such products on the nForce4 Ultra and even on the i915P chipset! For example, the PCI Express slot for a second graphics card on the Gigabyte GA-8I915P-SLI mainboard is made by joining lanes intended for independent PCI Express x1 slots, and the DFI LANPARTY UT Ultra-D uses the nForce4 Ultra chipset that doesn’t officially support SLI configurations. Thus, the range of chipsets that can support two graphics cards on NVIDIA’s chips doesn’t include only the nForce4 SLI and the i7525. In fact, the configuration “PCI Express x16 + PCI Express x2” is also possible, but the performance hit would be too obvious and this configuration would lack balance.
NVIDIA didn’t like the release of mainboards like the above-described, and the new versions of the ForceWare driver now block the SLI mode if the mainboard’s configuration doesn’t meet certain requirements. NVIDIA’s position is quite understandable: such systems have a low performance and this may affect negatively the reputation and popularity of the new technology.
Besides an appropriate mainboard, two identical graphics cards are necessary for the SLI mode. They can be either two GeForce 6800 Ultra/GT or two GeForce 6600 GT cards; a GeForce 6600 GT wouldn’t work in the SLI mode with a GeForce 6800 Ultra (or GT). In order to avoid compatibility-related problems NVIDIA recommends that you purchase cards from the same manufacturer for your SLI system.
Don’t forget that two graphics cards will inevitably call for more juice, so you may want to upgrade your current power supply, replacing it with a high-wattage unit with stable output voltages. For example, we powered our test system up with the help of a Cooler Master Real Power RS-450-ACLY unit and enjoyed perfect stability even with two GeForce 6800 Ultra cards inside. The system case should also be roomy enough and should have good ventilation, especially if you’re building a SLI configuration out of two top-end graphics cards.
The next section of this review describes the assembly and launch of a SLI-compatible system by the example of an ASUS A8N-SLI Deluxe mainboard and two pairs of GeForce 6800 Ultra and ASUS Extreme N6600 GT Series (GeForce 6600 GT) graphics cards. We will also examine these products closer in the meantime.
ASUSTeK Computer, the hardware manufacture of worldwide renown, offered us a kit consisting of two ASUS Extreme N6600 GT Series graphics cards and an ASUS A8N-SLI Deluxe mainboard (based on the NVIDIA nForce4 SLI chipset). We’ll talk about the graphics cards shortly, but now let’s scrutinize the mainboard, the basis of our SLI platform.
The A8N-SLI Deluxe mainboard came to our test lab in its retail package: a simple black box with the manufacturer’s logo, the product’s name, and the symbol of ASUS Proactive technology. There are no gaudy pictures or transparent windows, no advertisements. Everything’s official and stylish, as becomes a high-end device. It’s only on the left side of the box that we find some more text – it is describing the mainboard’s technical characteristics that are truly very impressive.
The word “Deluxe” in the mainboard’s name sounded much to us – ASUS’s Deluxe series products usually come with luxurious accessories. This case was not an exception. Here’s what we found inside the box:
The contents of this package can satisfy any user, but we think a product of the Deluxe class might have been equipped with aerodynamic ATA and FDD cables, which ASUS didn’t include. There is no wireless adapter among the accessories which is a prerogative of ASUS’s Wireless Edition mainboards.
We took the mainboard out of its antistatic bag to have a closer look:
The board itself looks gorgeous, being up to its status. ASUS lacquered the PCB deep brown and that looks real cool. The mainboard carries seven expansion slots: two PCI Express x16, two PCI Express x1, and three PCI slots. The DIMM slots are grouped in twos; the first slot of each memory channel is colored blue. The EZ Selector, a card that chooses the PCI Express mode, was already set to the Dual Card position on the mainboard, i.e. the default configuration is “PCI Express 8x + PCI Express 8x”. If you turn the EZ Selector by 180 degrees and plug it in again, the top slot becomes a PCI Express x16, and the bottom slot a PCI Express x1, because the nForce4 SLI chipset supports twenty PCI Express lanes.
Besides the four Serial ATA II ports and two ATA-133 channels, typical for nForce4 SLI-based mainboards, we have an additional SiI3114C controller from Silicon Image here. This chip procures four more Serial ATA-150 ports for the user. Thus, you can attach as many as eight Serial ATA devices to this mainboard – this seems a little excessive for today since optical drives with this interface are yet very rare. The mainboard fully supports nvRAID technology; you can build RAID arrays of levels 0, 1, 0+1, 5, 10 and JBOD.
The mainboard’s networking capabilities are courtesy of two Marvell chips. The Marvell Yukon 88E801 is a regular single-die Gigabit Ethernet controller that resides on the PCI bus; the Marvell Alaska 88E1111 chip is a physical-level (PHY) interface to the Gigabit Ethernet controller integrated into the chipset. ActiveArmor technology, the hardware firewall from NVIDIA, is fully supported by the nForce4 SLI.
The voice of the ASUS A8N-SLI Deluxe is the voice of the last-generation AC’97 ALC850 codec from Realtek. This is of course no SoundBlaster Live 24-bit that we saw on the MSI K8N Diamond mainboard, and even no High Definition Audio from Intel, but the quality and functionality of this codec should suffice for a majority of users, especially considering its ability to output digital audio through the coaxial and optical SPDIF outputs located on the mainboard’s connections panel.
The first PCI Express x16 slot is colored blue; the second (for the slave card) is painted black. These slots have special fixing locks. There are two PCI Express x1 slots in between, and you won’t be able to use one of them if you’ve got a GeForce 6800 Ultra or any other device with a dual-slot cooling system, installed. A Molex-type connector is located above the first slot; it is labeled as EZ_Plug and sends more power to the PCI Express bus in case two graphics cards are installed in the system. As you know, the maximum power this interface can provide is 75 watts, so additional power would be of much help with two cards. ASUS says EZ_Plug improves the system stability, and we agree wholeheartedly. There’s a red LED indicator near this power connector and it lights up if you’ve forgotten to attach the additional power. The second, green LED is placed under the PCI slots; it signals that the mainboard is receiving power.
The CPU power circuit is designed properly, although without a sophisticated cooling system. Its hottest components, i.e. the power transistors, are just equipped with a heatsink. Unfortunately many manufacturers forget about the cooling of these elements altogether, but these components are directly responsible for the well-being of your central processor. The CPU voltage regulator is three-channel on the ASUS A8N-SLI Deluxe.
The mainboard has five 3-pin connectors for attaching fans, but one of them is already occupied with the chipset fan, and one more is intended for the CPU cooler, so the user has to do with the remaining three connectors, two of which don’t support the speed monitoring feature. Q-Fan technology allows controlling the speed of the CPU cooler by setting the threshold temperatures at which it should stop, start and spin up to its maximum speed.
The mainboard’s connections panel is rather typical and has all the necessary ports, save for a COM port instead which there’s a FireWire connector. One COM port is available on the bracket for the back panel system case. This is in fact a right decision since FireWire peripherals are more popular nowadays than COM devices. The ASUS A8N-SLI Deluxe supports two FireWire ports; the second is available on a bracket. There are four USB 2.0 ports at the mainboard’s back panel, and four more can be attached to additional brackets. The nForce4 chipset supports ten USB 2.0 ports in total, and the remaining onboard USB headers can be used for a card-reader, for example, or you can output the ports to the system case’s front panel.
There’s nothing very interesting on the underside of the ASUS A8N-SLI Deluxe – only a metal plate that’s part of the CPU cooler’s fastening.
Although the PCB layout of this mainboard is overall good, we found some minor defects. For example, the placement of the PCI Express x1 slots isn’t very convenient: if two GeForce 6800 Ultra cards are installed, you unavoidably lose the access to one of these slots as well as to one PCI slot. You’ll only have two PCI slots left, and we wouldn’t recommend you to plug anything into the second PCI slot from top as this would surely block the airflow towards the cooling system of the lower graphics card. One free PCI slot – is it enough? It may be enough considering the high level of integration of this mainboard, but if, for example, you wanted to use a SLI graphics array, a TV-tuner and a high-end standalone audio card, you’d hardly succeed. Moreover, each long expansion card automatically blocks the bottom latches of all memory slots. It’s much better with two GeForce 6600 GT: you’ll have both PCI Express x1 and two PCI slots available.
The position of the chipset can’t be called appropriate, either. It is put between PCI Express x16 slots and if you install two GeForce 6800 Ultra or GT cards equipped with non-standard cooling systems, it’s possible the bottom card will press against the chipset cooler. The chipset’s heatsink also hinders your using the PCI Express x16 slot latches, especially if a GeForce 6800 Ultra or some other lengthy device is there. By the way, we didn’t like the chipset’s cooler much, either. It is a small heatsink with few ribs and a tiny fan. It couldn’t but provoke our worst apprehensions. Our doubts about its efficiency proved true later as the heatsink became very hot during our tests.
We didn’t check the problem of connecting all the power cables because of a simple reason that we’re assembling our test platforms without a system case to simplify and accelerate the process of changing the graphics cards. But we think there should be no problems here since the power connectors of the ASUS A8N-SLI Deluxe are located exactly like on the MSI K8N Diamond which we had no problems with (for more details about the MSI mainboard you can check out our article called MSI K8N Diamond (NVIDIA nForce4 SLI) Mainboard Review).
This mainboard refused to work with graphics cards on GPUs from ATI Technologies at first, but this problem was solved in BIOS 1002 Final, so that’s not critical anymore. But unfortunately we received this version of the BIOS only after we had done all our tests and written this review, so we measured the performance of ATI’s RADEON X cards on another platform.
The new mainboard from ASUS is a very well-made product on the whole, if you put up with the minor flaws in the PCB layout, but that’s actually a common problem of many, if not all, highly integrated mainboards on the nForce4 SLI chipset since it’s really hard to put so many variegated components within the confines of a PCB of the standard form-factor. The problem is not even in an inconveniently positioned expansion or memory slot, chipset and so on, but rather in the catastrophic lack of expansion slots if there are two graphics cards in the system. On the other hand, the ASUS A8N-SLI Deluxe has almost all the controllers you might want, and the components layout is made such as to give you at least one free PCI Express x1 and one PCI slot even with two installed GeForce 6800 Ultra cards.
Pros:
Cons:
We received a pair of ASUS Extreme N6600 GT graphics cards as an addition to the A8N-SLI Deluxe mainboard, the foundation of the SLI platform we’re assembling. Unlike the mainboard, the graphics cards came to us as they come to OEMs, i.e. packed in antistatic bags only. We needed these cards to investigate SLI technology, so it doesn’t matter much. Here’s how an ASUS Extreme N6600 GT graphics card looks like:
It differs from the classic GeForce 6600 GT with the deep blue color of the PCB only; the ASUS card is otherwise identical to the reference sample, except the design of the cooling system. The cooler from ASUS has the dimensions of the reference one, but that’s all there’s similar between the two. The heatsink here is a copper plate with ribs on its right. The fan made of translucent plastic has nine blades (against the reference fan’s seven blades), but the blades are thinner. Contrary to our expectations, there’s no highlighting of the fan’s blades.
This entire contraption is covered with a translucent aerodynamic casing so that the air stream from the fan went through the heatsink’s ribs and outside, also cooling the memory chips. In fact, we could say that this is a diminished and simplified version of the GeForce 6800 Ultra reference cooling system, but we’ve got an axial cooler rather than a blower here. This cooler is quiet – you don’t hear its noise against the noise from other PC components, even outside the system case.
Unlike many other manufacturers, ASUS put heatsinks on the memory chips on its GeForce 6600 GT. There two copper heatsinks each cools a pair of chips. No intricate fastenings here – the heatsinks are simply glued to the chips with thermal glue. The chip closest to the back of the PCB is additionally blown at by the GPU cooler. The ASUS Extreme N6600 GT is positioned as a device suitable for overclockers, so these heatsinks confirm this product positioning.
Especially for such users the GDDR3 chips have an access time of 1.6 nanoseconds rather than 2.0 nanoseconds as on a majority of other GeForce 6600 GT cards. We couldn’t check up the marking on the chips as their heatsinks were dead-glued to them, but our overclocking attempt bears indirect evidence to that: the memory worked stable at 600 (1200DDR) MHz. We also lifted the GPU frequency up to 600MHz, so this graphics card is true to the Extreme designation in its own name.
The 2D quality was up to the mark, too. The card yielded a crystal-sharp picture in all resolutions up to 1800x1440@75Hz inclusive.
We dealt with two OEM samples of the ASUS Extreme N6600 GT graphics card, but these cards will come to retail fully rigged out, with numerous accessories as you see with other products from ASUS. We’re left highly pleased with the Extreme N6600 GT as we just can’t find any obvious faults with this device.
NVIDIA Corporation was kind to offer us a couple of high-performance GeForce 6800 Ultra graphics cards for our tests:
Externally, the GeForce 6800 Ultra PCI Express differs from the GeForce 6800 GT in the design of the cooling system alone. It takes two slots and is analogous to the one we saw on the reference AGP-interfaced GeForce 6800 Ultra sample we described in our article devoted to the announcement of the NV40 GPU (see our article called NVIDIA GeForce 6800 Ultra and GeForce 6800: NV40 Enters the Scene). There’s a small, but important nuance, however. The GPU heatsink in the first revision of the cooler had been made of aluminum, but the company later changed the material for copper, to improve the cooling efficiency. In all probability the GeForce 6800 Ultra will use a copper heatsink henceforth.
Besides improving on the GPU cooling system, the shape of the heatsink that cools the components of the card’s power circuitry has been changed. It has become bigger, with more ribs. They must have thought the heatsink of the older type insufficiently good.
Launching a system with two installed GeForce 6800 Ultra graphics cards was not unlike launching a space shuttle – you shouldn’t be nearby at either of these events unless you want to lose your hearing. This pair of cards produced enormous noise, and what was particularly disgusting they wouldn’t slow down their fans after the OS was booted up, like GeForce 6800 Ultra/GT cards usually do! Of course we had to brace ourselves up and run our full cycle of tests under such conditions, but we don’t think home users are going to appreciate that. Our checking with the help of RivaTuner showed that the cards were working at 437/1130MHz frequencies (GPU/memory), and that was considerably higher than the normal frequencies of the GeForce 6800 Ultra graphics card.
We found out later that the problem with noise and frequencies only showed up on the nForce4 SLI platform. When installed separately into an i925-based system these graphics cards behaved quite normally, reducing their fan speeds at OS’s boot-up and working at the regular 400/1100MHz frequencies. At the moment of our tests the ASUS A8N-SLI Deluxe mainboard had had BIOS version 1003-001 Beta dated 30.11.2004, but the problem persisted even after we had updated the BIOS to version 1002 Final and installed the latest version of the ForceWare driver.
We are not sure what’s the root of the above-described problem, and we just hope end-users won’t meet it. We guess there would be few people who’d like the idea of wearing protective headsets before their computers.
This section of the review is about assembling a SLI-compatible computer. We will also describe step by step how to make it work correctly. Check up if you’ve got the following, first:
A comment about the last item in the list: NVIDIA says that the wattage of the power supply be at least 350 watts and recommends 500 watts and more for a pair of two GeForce 6800 Ultra. We used a 450W PSU from Cooler Master in our tests and had no stability-related problems. Now, let’s get to the assembly.
Suppose we’ve already loaded the CPU and memory slots of the mainboard with appropriate chip and modules and the mainboard itself is already mounted in the system case. We now want to enable the necessary PCI Express mode. In our case, the mode selector was set into the “Dual Card” position by default. This gave us the configuration “PCI Express x8 + x8”. And if the selector is set to the other mode, you should just turn it by 180 degrees.
With the A8N-SLI Deluxe, you push on the two locks at the sides of the slot and pull the selector card out. Then turn the selector, insert it back in the slot and press on it until the locks snap in place. The mainboard is ready to take in the graphics cards. Other manufacturers may have other selector devices or even use a jumper instead.
Now we must install two identical SLI-compatible graphics cards into the PCI Express x16 slots. Having installed them, link them through their SLI connectors using a special adapter and attach one of the PSU’s power plugs to the EZ_Plug connector (this last step may not be necessary on other mainboards, but we’re describing the installation procedure for the ASUS A8N-SLI Deluxe mainboard with all the particulars). Fasten the cards to the system case, but don’t forget about the special support that keeps the SLI adapter in place (this step is optional, too). Attach power to the graphics cards, if they require that. Note that a pair of GeForce 6800 Ultra cards needs four Molex connectors, so make sure beforehand your system case has them. Now you only have to plug the monitor’s cable to the graphics card in the top PCI Express x16 slot, and that’s all.
On the next start-up go into the BIOS Setup to check out the SLI Mode option. By default, it is set to “Auto” on the A8N-SLI Deluxe. That is, the mainboard itself determines the necessary mode, depending on the position of the mode selector. If you disable the automatic identification for some reason, make sure the option you select matches the mode set up by the selector.
When Windows XP boots up, you should install the required drivers, including nForce Unified Driver and ForceWare. You can download the latest official versions from the NVIDIA website. If all is done right, the NVIDIA Settings icon on the right of the Windows Taskbar will tell you that the system is ready for the SLI mode by showing a balloon with the text “SLI capable system” after the reboot.
Clicking on this message opens up the SLI multi-GPU dialog window:
You can now click the Enable SLI multi-GPU checkbox and reboot. The Show GPU load balancing option will display an indicator of the load distribution. This indicator is on by default, but you can disable it if you don’t like the green bar on the screen.
The SLI multi-GPU dialog window can be opened by clicking on the Advanced button of the Settings tab of the Display Properties window:
If you enable the multi-GPU mode, the driver will require that you reboot the computer. Agree to this proposal and after the reboot enter the driver settings window again and click the Show Advanced Settings option on the Performance & Quality Settings tab:
The SLI Rendering Mode radio button determines the behavior of the graphics card pair in case a game is launched which is missing in the database of the ForceWare driver. By default the button is set to “Single-GPU rendering”, so only one graphics card will be used if such an unknown game is started. If “Multi-GPU rendering” is selected, the system will try to use the multi-GPU mode even in unfamiliar applications. In this case you can get a performance boost or you can have image artifacts and/or system instability. It’s up to you which mode to use. Having chosen a mode, click the Apply button. The system is ready to run your games.
In the next section you will learn the configuration of our test platforms, the driver settings and the list of games we used for benchmarking.
We made use of three test platforms that were configured like follows:
NVIDIA nForce4 SLI platform:
PCI Express platform:
AGP platform:
Software:
We had to use the Intel Pentium 4 platform to test PCI Express graphics cards on ATI’s RADEON GPUs since the ASUS A8N-SLI Deluxe mainboard wouldn’t start up with them with the older version of its BIOS.
Since the NVIDIA nForce4 SLI platform and the testbed based on the VIA K8T800 Pro used the same processor (Athlon 64 4000+), we decided to compare the results of the SLI configurations with the results of single AGP 8x cards. We think this comparison is quite correct for today.
To avoid confusion, we mark the NVIDIA nForce4 SLI platform with an asterix (*) and the Intel 925X platform with a cross (†) in the diagrams.
Besides the GeForce 6800 Ultra and GeForce 6600 GT cards we united in the SLI mode, we had the following graphics cards in our tests:
AGP graphics cards:
PCI Express graphics cards:
The Catalyst A.I. option was set to “Standard”, and the Mipmap Detail Level option to “Quality”. We disabled VSync in the drivers for all the participating cards. We also turned off the ForceWare optimizations save for the Anisotropic mip filter optimization, using the Quality mode.
We turned on full-screen antialiasing and anisotropic filtering from the game’s own menu. Otherwise we forced the necessary mode from the driver. The maximum graphics quality settings were selected in each game, the same for graphics cards on ATI’s and NVIDIA’s GPUs.
We enabled the Multi-GPU Rendering mode for the SLI platform, i.e. we forced the SLI mode even for games the ForceWare driver doesn’t officially support. Thus we checked out how efficient SLI technology is in such cases, if it works at all.
At first we wanted to use the same games as in our mega-roundup of the graphics hardware made in 2004, but we were stopped at the capriciousness of the NVIDIA SLI platform. Particularly, the games from the list below either wouldn’t start at all or were unstable, hanging up regularly:
Well, you can of course launch each of these games on a GeForce6 SLI system, but using only one of the graphics cards to do the rendering. It’s not exactly what you’d expect from a $500 graphics subsystem?
Some other games would have no speed bonus on the SLI platform or would even run slower on it than on a single graphics card, but if the game and the system remained stable, we included such games and results into our review. After we had thrown away unstable or unlaunchable games, we got the following list:
First-Person 3D Shooters:
Third-Person 3D Shooters:
Simulators:
Strategy Games:
Semi-Synthetic Benchmarks:
Synthetic Benchmarks:
You don’t see Counter-Strike: Source and Vampires the Masquerade: Bloodlines due to their very low speed on the PCI Express platform.
Against all expectations, we witnessed few image-quality-related problems in the SLI mode; in fact we only saw them in the flight simulator IL-2: Sturmovik.
As you can see, the problem is evidently connected with the pixel shaders the water surface is rendered with. This problem is likely to be corrected in a ForceWare update or a patch from the game developer as IL-2 is among those games where the advantages of SLI technology are most conspicuous. Without doubt, the stability issues in some games will also be seen to very soon, but at the moment of our tests the stability of the NVIDIA SLI platform was rather far from perfect.
NVIDIA’s multi-GPU technology wouldn’t work with this game. No, the system was perfectly stable and we had no image-quality-related problems, but we had no performance gains, either. Moreover, we didn’t even see the Load Balancing indicator, although we had enabled it in the driver.
It’s the same in the “eye candy” mode: no profit from the SLI mode whatever.
NVIDIA’s SLI did well in Doom 3, although its advantage is close to nothing in low resolutions: the scene recorded on the Hellhole level contains a battle with numerous monsters and the speed here mostly depends on the CPU characteristics rather than on the graphics subsystem. You can see the positive SLI effect in 1280x1024 where the GeForce 6600 GT duo performs like the single GeForce 6800 Ultra. The system with two GeForce 6800 Ultra showed its best in the highest resolution, but the performance gain was not twofold.
The efficiency of NVIDIA’s multi-GPU rendering grows up as we enable such eye-pleasing features as full-screen antialiasing and anisotropic filtering. The limiting influence of the CPU is felt less here, as the graphics subsystem bears a higher load. The GeForce 6600 GT couple evidently felt the lack of the graphics memory – the memory of the graphics cards in a SLI configuration has identical contents, i.e. its total amount remains the same (128 megabytes in this case) rather than doubles. We tested this game with the highest graphics quality settings, so the relatively low results of the GeForce 6600 GT SLI must be due to the low memory amount. The GeForce 6800 Ultra SLI looks better – the multi-GPU mode provides a performance boost up to 75 percent in high resolutions.
The d3dm4 scene is simpler: there are no monsters here, but the player is shooting from a plasma rifle or a rocket launcher from time to time. The SLI effect is noticeable even in the lowest resolution and is growing up to 60-65 percent in 1600x1200.
Again the lack of the memory affects the results of the GeForce 6600 GT SLI in the “eye candy” mode, so the SLI effect amounts to 60 percent in 1600x1200 only, while the pair of GeForce 6800 Ultra cards are 80-95 percent faster than the single card. We want to remind you that we tested the single GeForce 6800 Ultra on the VIA K8T800 Pro platform, so we should make allowances for the difference between the platforms when comparing the results.
The performance of the system with two GeForce 6800 Ultra cards is limited by the CPU in the first two resolutions of this game, but for the GeForce 6600 GT SLI configuration the performance gain is from 20 to 40 percent. In 1600x1200 the pair of GeForce 6800 Ultra enjoys a speed boost thanks to SLI, but the value of the gain is far from twofold because Halo is too CPU-dependent.
Note, however, that it’s only through the SLI effect that the GeForces can rival the ATI RADEON X850 XT Platinum Edition in this game.
NVIDIA’s multi-GPU technology wouldn’t work with Painkiller. We had no problems with the system stability or image quality, but we had no benefits from it, either. A pair of two cards would be as fast as a single card of the same type.
It’s the same situation in the “eye candy” mode, but in 1600x1200 the GeForce 6600 GT SLI configuration suddenly had a twofold advantage over the single GeForce 6600 GT which we hadn’t seen in lower resolutions.
SLI technology brings its positive effect to Half-Life 2, too. This effect is less conspicuous in low resolutions where the load on the graphics subsystem is low. In high resolutions, however, the performance in the SLI mode is 50-60 percent higher than that of the single card.
NVIDIA’s multi-GPU technology has an even bigger effect in the “eye candy” mode – up to 80 percent. Yet the SLI configuration has a really big advantage over the high-performance solutions from ATI Technologies in 1600x1200 resolution only.
The second scene from Half-Life 2 that we use in our tests depends more on the CPU speed than the first one, so it’s harder to see the SLI effect here, at least in the “pure speed” mode.
SLI technology’s speed bonus is evident in the “eye candy” mode, in 1280x1024 resolution and up, especially with the system consisting of two GeForce 6600 GT cards. The pair of GeForce 6800 Ultra cards outperforms the RADEON X850 XT PE and X800 XTPE in 1600x1200 resolution only, but NVIDIA’s SLI seems to be at its best under such hardest conditions.
No performance benefits here. But there are no stability or image-quality problems, at least!
The multi-GPU technology doesn’t work in Hitman: Contracts. Although the SLI platform is somewhat faster, this must be due to the differences in the configuration and architecture of the test platforms on NVIDIA nForce4 SLI and VIA K8T800 PRO chipsets.
Although the SLI platform produced a correct picture in this game, its speed was slower than of the single card. The system with two GeForce 6800 Ultra chips had a very low frame rate in 1600x1200, but this may be the outcome of our testing all third-person shooters manually, with the FRAPS utility. There’s inherent inaccuracy in this test method that can’t be avoided.
Unlike the three previous games, Deus Ex: Invisible War not just runs smoothly on the SLI platform, but runs up to 40 percent faster on the two GeForce 6600 GT cards!
IL-2 Sturmovik would be an excellent demonstration of the benefits of SLI technology, but to our regret it is in this game that we encountered troubles with the quality of the picture. The water surface suffers the most, so you’d hardly like to play that, even though you can get an almost twofold performance gain by enabling the multi-GPU mode. This problem will be hopefully corrected soon.
The problems with the image quality put aside, the SLI effect is simply astonishing. The SLI platform ensures a comfortable frame rate even in 1600x1200 with enabled full-screen antialiasing and anisotropic filtering! No single card, including the ATI RADEON X850 XT PE, can match the speed of the multi-GPU technology from NVIDIA in IL-2.
This flight simulator depends heavily on the CPU, so we can’t tell if SLI is any good here, at least in the “pure speed” mode.
The results of the “eye candy” mode have something to tell us: in 1280x1024 and in 1600x1200 the SLI mode works correctly and brings some speed advantages.
Alas, FlatOut can’t say anything interesting to us in the “pure speed” mode, since this game has a frame-rate limiter set to 100fps. Almost all modern graphics cards can easily reach this barrier.
The “eye candy” mode helps to see what was hidden in the “pure speed” one: the SLI speed bonus amounts up to 50 percent here.
Here’s one of those cases when the multi-GPU rendering mode only worsens the performance of the system.
It’s different in the “eye candy” mode: there’s a nice performance gain from the multi-GPU rendering in high resolutions.
The GeForce 6600 GT AGP graphics card behaves queerly here – it’s far behind its PCI Express analog.
Dawn of War is a relatively new game which has already become widely popular among the fans of real-time strategies. NVIDIA’s SLI works well here, providing a performance gain from 30 to 50 percent. This helps the NVIDIA team to win this test from ATI’s RADEON X850 XT PE and X800 XT PE.
The multi-GPU mode yields excellent results in the “eye candy” mode, too, but only with the two GeForce 6800 Ultra. As for the cheaper SLI configuration (made of two GeForce 6600 GT cards), it is slower than the single GeForce 6800 Ultra, probably due to the same reason as in some earlier tests – the insufficient amount of the graphics memory.
The new technology boasts a nearly maximum efficiency in this highly demanding game. Perimeter has always favored ATI’s GPUs, but now they cannot compete with the tremendous power of the two NV45 processors.
It’s the same in the “eye candy” mode with enabled full-screen antialiasing and anisotropic filtering, but the GeForce 6600 GT SLI pair once again is limited by the amount of the graphics memory. Besides that, two 128-bit busses do not equal one 256-bit bus: their total efficiency will surely be lower for the same very reason for which two SLI graphics cards are not exactly two times faster than a single card.
The SLI mode gives no profit in Rome: Total War. In the “pure speed” mode this game, like a majority of real-time strategies, depends on the CPU performance rather than on the graphics subsystem.
The game is rather wayward as a benchmark, but we can see some traces of the SLI effect in 1280x1024 – the performance of the two GeForce 6600 GT is evidently higher than of the single card.
SLI technology provokes a horrible performance hit in this test, although there are no problems about the quality of the picture.
The multi-GPU technology from NVIDIA works well in this test, although the performance gain is far from the theoretical maximum. The GeForce 6800 Ultra duo even outperforms the top-end graphics cards from ATI.
ATI’s graphics cards have traditionally showed their best in the “eye candy” mode. They do so in Aquamark3, too, but now NVIDIA’s GPUs in the SLI mode can deliver the same performance. The value of the performance gain is about 50 percent for the GeForce 6600 GT SLI and about 40-45 percent for the GeForce 6800 Ultra SLI.
Quite expectedly SLI technology is perfectly compatible with the benchmarks from Futuremark, and the results are another testimony to its efficiency: the SLI platform overcame the next landmark, a total score of 20,000 points. The GeForce 6600 GT pair also did excellently, outperforming the single GeForce 6800 Ultra.
The first test is simple, so we can’t expect a big effect from enabled SLI, but it is present, especially in high resolutions. The pair of two GeForce 6600 GT cards profits the most from SLI, achieving a performance gain of 30-35 percent. The GeForce 6800 Ultra SLI configuration is 15-20 percent faster than the single card. The multi-GPU mode helps NVIDIA’s GPUs to outperform their rivals from ATI Technologies.
The advantages of NVIDIA SLI technology are more conspicuous in the “eye candy” mode – the performance gain is about 75 percent for the GeForce 6800 Ultra SLI in 1600x1200 resolution, and about 100 percent for the GeForce 6600 GT SLI!
NVIDIA’s innovation looks even more impressive in the second test: the cards working in the SLI configurations deliver almost two times the performance of the respective single cards – in all resolutions!
The same goes for the “eye candy” mode, but the SLI configuration made of two GeForce 6600 GT cards again lacks graphics memory and falls behind the single GeForce 6800 Ultra.
The bonus from the enabled SLI mode is smaller in 3DMark03’s third game test, but is still close to the theoretical maximum.
Unlike in the second test, the GeForce 6600 GT SLI configuration is no slower than the GeForce 6800 GT in the “eye candy” mode. The load on the graphics memory subsystem must be rather low here.
ATI’s graphics cards have always been leaders in the fourth test, but this time the GeForce 6600 GT SLI matches them, and the SLI pair of two GeForce 6800 Ultra cards wins this test. The efficiency of NVIDIA’s multi-processor technology is close to the theoretical maximum with the latter configuration.
SLI is beyond praise in the “eye candy” mode, too. But again we see the GeForce 6600 GT pair hamstringed by the lack of the graphics memory in high resolutions. That’s why this configuration is slower than the RADEON X800 XT PE in display modes above 1024x768. The GeForce 6800 Ultra SLI takes its deserved first place.
Joining two GeForce 6600 GT cards in a SLI configuration leads to a twofold performance gain – they can even contend with the RADEON X850 XT PE/X800 XT PE. The efficiency of the GeForce 6800 Ultra SLI is somewhat lower, about 85-90 percent, but that’s a good result anyway.
The performance gain is about 90-100 percent with both SLI configurations in the first game test. That’s why the more expensive configuration of two GeForce 6800 Ultra cards wins this test, while the less expensive SLI system of two GeForce 6600 GT keeps even with ATI’s top-performance devices.
The GeForce 6600 GT SLI system refused to run this test in high resolutions with enabled full-screen antialiasing and anisotropic filtering due to the lack of the graphics memory. Yet it was a little behind the single GeForce 6800 Ultra even in 1024x768 – that’s another proof of the necessity of big amounts of graphics memory in modern applications. As for the GeForce 6800 Ultra SLI platform, its efficiency in the “eye candy” mode is going down as the resolution grows up. In 1600x1200 this system is just a tiny step ahead of the RADEON X850 XT Platinum Edition.
The new technology looks better in the second test where the speed of the coupled GeForce 6600 GT cards is no worse than that of the RADEON X850 XT PE or X800 XT PE. The two GeForce 6800 Ultra in the SLI mode are exactly two times faster than the single card of the same type.
With enabled full-screen antialiasing and anisotropic filtering we see the same situation as in the first test, but the performance hit the multi-GPU system experience in high resolutions is smaller, while their efficiency is overall higher.
The third test can be regarded as the most difficult one; it uses sophisticated pixel shaders. Graphics cards on GPUs from ATI technologies have always been winners here, but NVIDIA’s implementation of the multiple GPU concept breaks this tradition: you see that the GeForce 6800 Ultra SLI configuration is 50 percent faster than the RADEON X850 XT Platinum Edition. The SLI due of two GeForce 6600 GT cards are very fast, too. Their performance is higher than that of the single GeForce 6800 Ultra and about the same level with the RADEON X800 XT Platinum Edition.
The GeForce 6600 GT SLI isn’t as confident in the “eye candy” mode as in the “pure speed” one and is behind the single GeForce 6800 Ultra. The two GeForce 6800 Ultra once again have a twofold advantage over the single card, though.
On the whole, the high scores of the SLI platform in both versions of 3DMark are a good reflection of its efficiency in 3DMark’s tests. NVIDIA’s multi-GPU technology sometimes reaches its theoretical efficiency there.
So after a long interval we see another attempt of the promotion of a technology that joins several graphics processors to increase the performance of the graphics subsystem of the computer. We have tested the technology developed and implemented by NVIDIA Corporation, and our impressions about it are rather ambiguous.
Let’s consider it an established fact that the SLI mode does produce a considerable performance gain, up to 100 percent, or twofold, in some cases. NVIDIA’s developer team did a very good job and had brought all possible latencies and time losses associated with synchronizing two graphics card to the lowest minimum possible.
That said, we could agree with NVIDIA’s claims about a super-high efficiency of SLI technology if it were not for one problem. SLI does not work with all games. In many gaming applications there was no performance gain whatever or there even was a negative effect from enabled SLI. In some cases enabling SLI made the system unstable, or a game just would not start up.
Another drawback of SLI technology is its dependence on the optimizations for each particular game in the ForceWare driver. If the game is optimized for SLI, the performance gain can be huge, but if the game is not in the database, there’s a high probability of your getting a very small speed bonus (far from the promised 70-90 percent) or none at all. In other words, the multiple GPU technology from NVIDIA has a high efficiency, but lacks the universality peculiar to the Video Array system from Alienware, for example.
NVIDIA’s innovation also proved to be most fastidious about the installed cards – only two identical cards from the same manufacturer would work together. We couldn’t make two GeForce 6600 GT from different makers work in the SLI mode, despite their having identical characteristics and the necessary connector.
This is not to say that NVIDIA’s SLI is bad or defective. That’s just “childhood diseases” that every new technology suffers from. Well, even the nForce4 SLI platform itself couldn’t boast maturity at the moment of our tests – the ASUS mainboard worked with a beta BIOS, and with a beta version of NVIDIA’s drivers for the new chipset. This couldn’t but tell (and it did tell) on the stability of the new platform. Beyond doubt many drawbacks now peculiar to NVIDIA SLI will be eliminated sooner or later. There will be more supported games and the Multi-GPU rendering mode will work more correctly.
There’s one more important aspect about computer systems with several graphics processors/cards which is the main hindrance for such systems to become widespread and popular among ordinary users. That’s price, of course. Let’s calculate what money the highest-performing version of the nForce4 SLI platform may cost you. Suppose you already have a CPU for Socket 939 and two PC3200 memory modules, and a high-wattage PSU. You need a SLI-compatible mainboard and two SLI-specified graphics cards on NVIDIA’s GPUs.
The ASUS A8N-SLI (ordinary or Deluxe version) will cost you about $200. A PCI Express version of the GeForce 6800 Ultra is priced at $500. A pair of less expensive GeForce 6800 GT cards is going to cost $400 per item, according to NVIDIA. So you’re going to have to pay as much as $800-1000 for the graphics cards only. And these are the best prices since many resellers are selling the GeForce 6800 Ultra for a price high above the recommended one. The total cost of such a computer would be approaching the $4000-5000 mark, which is too high for a home gaming system, even of a highest class. Anyway, this platform will find its client as people who are not limited in money can go for such a purchase. If some game doesn’t work in the SLI mode, you can always switch to the Single-GPU rendering mode and have the performance of a single GeForce 6800 Ultra.
It’s different with the two GeForce 6600 GT. A pair of these cards will cost about $400, without the cost of the mainboard. This seems to be quite affordable for many gamers. At first sight one may think that two GeForce 6600 GT cards are equivalent of the GeForce 6800 Ultra and the RADEON X850 XT Platinum Edition in terms of the number of pixel and vertex processors as well as the amount of graphics memory. But simple arithmetic doesn’t work here – one and one is not necessarily two.
The problem is that besides the synchronization overhead and the lack of support of some games, the amount of graphics memory of two GeForce 6600 GT cards will remain 128MB, rather than 256MB, since the cards in the SLI mode have identical memory contents. This memory size isn’t appropriate for a high-performance solution intended for high resolutions and full-screen antialiasing. The results of our tests confirm this point: the GeForce 6600 GT SLI configuration would be slower in some cases than the single GeForce 6800 Ultra or RADEON X850 XT Platinum Edition or X800 XT with enabled FSAA and aniso-filtering. High resolutions became inaccessible at all in Rome: Total War and 3DMark05 due to the lack of the graphics memory.
That said, we don’t think it’s profitable to purchase two GeForce 6600 GT cards to unite them into a SLI graphics array. You can buy a GeForce 6800 GT or a RADEON X850 XT for about the same money and have the same performance, but without those problems about the compatibility and the amount of the graphics memory.
It’s too early yet to make a final judgment about the success of NVIDIA SLI technology, but it is not a failure, that’s certain. When there were no compatibility and stability problems the technology boasted a very high efficiency: adding a second graphics card and enabling the SLI mode we could lift the performance 1.5-2 times, while the analogous technology from Alienware could give you only 40-60 percent relative to the single card.
We guess it is the GeForce 6600 GT SLI configuration that’s not very appropriate. A platform with two GeForce 6800 Ultra may be demanded by wealthy PC enthusiasts, but the “inexpensive” realization of NVIDIA’s multi-GPU technology proved to be not so cheap in practice: its price is comparable to the price of single top-end cards, and the latter have almost the same performance, but better stability and compatibility with the existing gaming software.
You maybe know that NVIDIA is not alone in its interest in multi-GPU solutions – this company is just the first to implement the concept in silicon. ATI Technologies is also planning to unveil its multi-GPU technology soon. It is already given a name – ATI Multi Rendering or AMR. This technology is expected to be more flexible than NVIDIA SLI: it will allow using different graphics cards from ATI and will free the user from the additional adapter – the information necessary to synchronize the cards will be transferred through the PCI Express bus. S3 Graphics is also preparing an analogous solution, so there’s going to be a wider choice of multi-GPU platforms very soon.
The results of the poll we conducted on our site (see this page for more details) suggested that NVIDIA SLI was not going to become a sensation among PC gamers. Only 5 percent of the participants said they wished to buy a SLI kit of a mainboard and two graphics cards. 15 percent more voted in favor of purchasing one card and then, some time later, another one. On the other hand, 45 percent of users are considering the possibility of buying a system with some version of the multi-GPU concept. That’s why we don’t think the new incarnation of the idea of multi-GPU graphics hardware is going to be a complete failure.