by Alexey Stepin , Anton Shilov
09/26/2005 | 08:06 PM
Consumer market oriented multi-processor graphics technologies have gone a difficult and thorny path. Numerous attempts have been made to promote such solutions since 1997 when the late 3dfx proposed to join two Voodoo2 graphics cards to get more performance in 3D applications. ATI Technologies then released its Fury MAXX graphics card with two Rage Fury chips and XGI developed an impressive, yet unviable dual-processor Volari Duo.
Neither of the mentioned solutions ever really took off. They were all ruined by their high price combined with various architectural and software problems. To be specific, ATI could not make the AFR rendering mode work right on the software level, while XGI’s solution was based on two low-performance Volari GPUs linked with a narrow bus. One way or another, multi-GPU solutions left the scene until 2004 when NVIDIA decided to revive the concept that seemed to have been forever buried. We dedicated an in-depth review to NVIDIA’s SLI technology (for details see our article called NVIDIA Multi-GPU SLI Technology: New Approach to Old Ideas) and thought it promising enough, despite some drawbacks like its dependence on the software.
We were right in our suppositions. SLI-supporting systems have become quite popular among PC enthusiasts who want to have the maximum gaming performance whatever the cost. The technology has also matured since its release, getting rid of some “growth diseases” and transforming into a finished and trusted solution. On the other hand, SLI proved to be less efficient when applied to mainstream graphics cards. The possible compatibility problems with some games and the necessity to buy a special SLI-ready mainboard made such SLI configurations less appealing than the purchase of a single high-end card.
Thus, NVIDIA’s SLI got firmly established in the top-end market sector and its position was further strengthened with the arrival of the new generation of graphics cards from NVIDIA. Two GeForce 7800 GTX cards in a SLI system set new performance records (for details see our article called NVIDIA GeForce 7800 GTX: Monstrous Gaming Performance Unleashed). Although the market of multi-GPU graphics solutions is rather narrow and is in fact limited to high-end products, NVIDIA’s SLI was impressively successful there, with about 2 million nForce4 SLI chipsets already sold. Today, the market of multi-GPU-compliant chipsets almost wholly belongs to NVIDIA. Has ATI any chance to get a slice of that pie? To do so, the company must offer a technology which is at least no worse than NVIDIA’s SLI in every aspect. Can CrossFire be this technology? That’s exactly the question we are going to answer with this review.
Before the release of NVIDIA’s SLI, the RADEON X800 XT Platinum Edition graphics card from ATI Technologies was the performance leader in the market of consumer graphics solutions. It surpassed the GeForce 6800 Ultra due to its higher clock rates, more efficient memory controller and increased speed at processing version 2.0 pixel shaders. Not supporting Shader Model 3.0 and OpenEXR HDR, the card was functionally inferior to NV45-based devices, but its performance was really higher in a majority of applications. The arrival of a technology that could join two NVIDIA GeForce 6800 Ultra or GT graphics cards into a single, even though very expensive, graphics subsystem dethroned the RADEON X800 XT Platinum Edition immediately, stripping ATI of the crown of the manufacturer of the fastest graphics solution. Even the RADEON X850 XT Platinum Edition released a little later couldn’t wrest the crown from SLI configurations based on NVIDIA GeForce 6800 Ultra and GT.
Being the main market opponent to NVIDIA, ATI Technologies couldn’t let this just pass. Rumors began to spread out immediately after the release of NVIDIA’s SLI that ATI was working on its own project called AMR – ATI Multi Rendering. ATI should be given credit as an experienced developer of multi-chip graphics solutions. At least, GPUs from this company work in professional simulation & visualization systems manufactured by Evans & Sutherland since 2001 and in SGI’s systems since 2003. ATI’s technology was expected to be free from the main downsides of NVIDIA’s SLI – to be more configuration-flexible by allowing to use different graphics cards, to be independent from the driver’s support, to synchronize the two GPUs via the PCI Express bus thus making a connecting adapter between the cards unnecessary.
The reality turned to be different, as usual. ATI officially announced its CrossFire technology on May 30, 2005. The technology lacked the expected flexibility of configuration. CrossFire-compatible cards are divided into Masters and Slaves and these are to be joined by means of a special external link. More expectedly, it turned out two different graphics cards (for example, a RADEON X800 and a RADEON X700) could not be used in a CrossFire system. Well, it would be very hard to ensure stable operation of such a system while the performance gain would be very small or even negative due to the problem of load distribution between two different GPUs. Master cards – equipped with a special frame-combining chip – were announced for RADEON X800 and RADEON X850 only.
Alas, CrossFire has only remained on paper ever since its announcement on May, 30. Real products – mainboards and graphics cards – were expected in June and July of this year, but after numerous delays due to some technological problems they still did not appear. Moreover, CrossFire took so long to reach the market that NVIDIA managed to release its new-generation graphics processor G70 in the meanwhile. G70-based graphics cards soon came to the shops and thus devaluated SLI configurations with two GeForce 6800 Ultra. As a result, the perspectives of CrossFire grew less certain, at least in its current, RADEON X850/X800-based incarnation.
With all the above-mentioned problems, CrossFire was still a very interesting technology, combining a number of original engineering solutions which we described in detail in our preview (for details see our article called ATI Crosses the Swords: Multi-GPU CrossFire Technology Previewed ). In brief, the main advantage of CrossFire over SLI is its theoretical capability to ensure the maximum performance gain in any fame, irrespective of its version and the drivers. This is achieved by means of a special chipset (Compositing Engine) that “seams together” the parts of the image rendered by different graphics processors. The Slave card sends the rendered part of the image via a special Y-shaped cable to the Master where the Compositing Engine combines it with the data from the Master GPU. The result is then sent to the monitor.
A CrossFire system can work in three modes that differ in how the load is distributed between the GPUs. The Scissor mode works in a similar way as the SMR mode NVIDIA implemented in its SLI technology.
The point of this mode is two split the frame in two parts each of which is processed on a separate GPU. For better balance, the size of the parts may vary dynamically. The frame is analyzed and divided by the graphics card’s driver. The downside of this mode is that the geometrical performance of the CrossFire system equals that of the single card because each card of the system processes the geometrical data for the entire frame. The Scissors mode can be used successfully with Direct3D as well as OpenGL applications.
The Alternate Frame Rendering (AFR) mode, originally introduced by ATI in its dual-chip Fury MAXX graphics card, works differently. In this mode the graphics processors process and output alternating frames, one after another.
Each GPU processes the geometrical information for its frame only which makes this mode more promising in terms of performance gains.
The SuperTiling mode is a completely new principle of distributing the load among several GPUs, developed at ATI Technologies. In this mode the image is divided into squares (tiles), 32x32 pixels each, following the checkerboard pattern:
One half of the tiles are processed by the Master card and the other half is the Slave’s responsibility. Since the size of each square is small, both graphics processors receive almost the same load automatically, without any intervention on the part of the driver. This mode works only with Direct3D applications and each GPU has to process the geometry of the entire frame. Today’s games, however, are generally satisfied with the vertex processor performance of modern graphics cards. For example, the 6 vertex processors of the RADEON X850 XT clocked at 520MHz frequency are quite enough for the lack of geometry performance scaling not to become a bottleneck of a CrossFire system.
The rendering mode of a CrossFire system is selected automatically, depending on the Catalyst A.I. option. If this option is enabled, the Scissors or SuperTiling mode is used depending on the API. The AFR mode is only enabled for applications that have the appropriate profiles written into the Catalyst driver. If the Catalyst A.I. option is turned off, the CrossFire system uses SuperTiling for Direct3D applications and Scissors in all other cases.
We mentioned three CrossFire modes above, but there is a fourth one. The Super AA mode differs from the other three in improving the image quality rather than increasing the performance. A CrossFire system makes 8x, 10x, 12x and 14x full-screen antialiasing modes available.
The 8x and 12x modes are improved versions of the 4x and 6x modes, respectively. In this case each graphics card works in FSAA 4x or 6x mode, but the samples at multi-sampling are taken from different locations. The quality of the subsequent merger of the two resulting frames is equivalent to MSAA 8x/12x. The 10x and 14x modes are hybrid ones, combining 2x super-sampling with 4x/6x multi-sampling for a near ideal quality of small image details and smoothing out transparent textures. The number of texture samples is doubled in the Super AA mode, so the anisotropic filtering range is doubled, too. In other words, AF 32x mode becomes available. Super AA is unavailable if you use one of the three performance-boosting rendering modes, but you can use the ordinary 2x/4x/6x FSAA with them.
As mentioned above, the information from the Slave card is transferred to the Master via an external Y-shaped cable in the DVI format. This is a snapshot of the cable:
The multi-pin DMS-59 connector is attached to a special connector on the PCB of the Master card, the male DVI connector – to the DVI connector of the Slave card and the female DVI-I connector to the DVI output of your TFT monitor or, via an ordinary DVI-I → D-Sub adapter, to any CRT monitor.
To implement different rendering methods in one graphics subsystem ATI had to use a programmable FPGA matrix, capable of changing its structure depending on the task at hand, as a chip that combines the parts of the image. The employed Xilinx XC3S400 chip belongs to the Spartan-3 series, consists of 400 thousand transistors, and has 8064 logical cells and 43 kilobytes of integrated memory.
To translate the resulting information into a FPGA-matrix-friendly form, a 165MHz TFP401A chip from Texas Instruments is used. The frame can be outputted to an analog monitor through an ADV7123 chip (a three-channel 10-bit 330MHz RAMDAC) or to a DVI-interfaced LCD panel through an ordinary TMDS transmitter SiI1162.
The single-link receiver from TI is actually a weak link since its operational frequency is only 165MHz, so the maximum supported resolution is UXGA (1600x1200). The use of this receiver is necessary because the Slave card has a 165MHz transmitter with the same limitation. This doesn’t seem to be a big problem as a majority of monitors in use simply do not support higher resolutions, but there’s an unpleasant surprise for any CrossFire user.
As you probably know, the required pass-band in megahertz is calculated by the formula R = (1.4*X*Y*V)/1000000 where 1.4 is a device lag compensation coefficient, X and Y are the horizontal and vertical resolutions, and V is the vertical sync frequency. You can solve this simple equation to see that a bandwidth of 165MHz is only sufficient to output a 1600x1200 picture at 60Hz frequency. A 75Hz refresh rate would require a bandwidth of over 200MHz; and an 85Hz refresh rate would require a bandwidth of about 230MHz. It turns out that CrossFire cannot do what all multi-GPU technologies are expected to do, i.e. to work in high resolutions! Yes, 1600x1200 is formally supported, but it is absolutely unacceptable to have a refresh rate of 60Hz. Talking about higher resolutions, for example 1920x1080, a bandwidth of 175Hz is required even for a 60Hz refresh rate. So, people who have a HDTV 1080i display device or a big CRT monitor may feel disappointed. Of course, LCD monitors or big-diagonal TV-sets will be able to use all the advantages of CrossFire, but not all gamers are yet satisfied with the response time of typical LCD monitors, so the limitations imposed on the display resolution is quite annoying.
A majority of TMDS receivers and transmitters work at 165MHz frequency and an additional receiver is usually employed to expand the capabilities of the main one (the so-called dual-link design). But an additional receiver would make the wiring of RADEON CrossFire Edition cards even more complex than it is now, so ATI must have decided to use only one chip. It is also logical because a different design of the Master card would make a Slave card, already purchased by the user, unsuitable for CrossFire.
Well, Silicon Image’s product range already includes a 225MHz SiI1171 receiver capable of working with resolutions up to 2048x1536 and CrossFire will probably use this chip in the future (especially if we recalled that the RADEON X1800 supports two dual-link DVI outputs), but so far the capabilities of ATI’s multi-GPU technology do not go beyond 1600x1200@60Hz. It is one of the main disadvantages of this technology and the price for the lack of an internal bridge between the two cards.
Yet another drawback of CrossFire is its configuration inflexibility. You need a Master card equipped with a Compositing Engine to build a CrossFire system. And if you own a 12-pipelined or an 8-pipelined RADEON X850/X800, the Master card will avoid misbalance by automatically disabling the “extra” pipelines and make the SuperTiling mode unavailable. The resulting efficiency of such a system is much lower than when two 16-pipelined RADEONs are in use. But even in the last case, if the frequencies of the two cards differ greatly (like those of the RADEON X850 XT CrossFire Edition and the RADEON X800 XL), the Master automatically drops its frequencies down with the ensuing performance drop.
CrossFire will probably appeal the most to users who already have a RADEON X850 XT/XT Platinum Edition/RADEON X800 XT/X800 XL. They just have to buy an appropriate Master (RADEON X850 XT CrossFire Edition, RADEON X800 XL CrossFire Edition or RADEON X800 CrossFire Edition). Other users are unlikely to get interested in the current incarnation of CrossFire because ATI’s new generation of graphics cards is expected soon and because NVIDIA’s GeForce 7800 GT/GTX graphics cards with comparable performance, better functionality, capable of working in multi-GPU configurations and not limited to 1600x1200@60Hz resolution, have been freely available since their announcement.
It is only at the end of September 2005 that ATI at last solved all the problems with its new technology, so the 26-th of September should be considered the real birthday of CrossFire. As we said in our preview of this technology, the leading graphics developers are currently of opinion that the best results can only be achieved on a specially prepared platform. In other words, we should talk the platform at large rather than graphics card with its accessories when it comes to multi-chip technologies.
For example, NVIDIA’s SLI works the best with NVIDIA’s nForce4 SLI chipset, even though theoretically it can work on any other platform equipped with two PCI Express x16 slots. The same it true for CrossFire. It requires a special mainboard based on either of the two versions of the RADEON XPRESS 200 CrossFire Edition chipset: RD480 (for AMD) and RD400 (for Intel). Chipsets from other manufacturers will get certified, too, but so far these are the only CrossFire-compatible chipsets available. Let’s see what they are right now.
The RADEON XPRESS 200 CrossFire Edition chipsets have a classic two-chip architecture and support all modern processors from Intel and AMD. The RD400 North Bridge (for the LGA775 platform) supports a 1066MHz FSB and has a memory controller compatible with DDR400 (PC3200) as well as with DDR2-667. The R480 is simpler in design since AMD’s modern processors have an integrated memory controller. This chipset is in fact a PCI Express controller compatible with AMD’s processors that use a 1GHz HyperTransport bus.
The ATI chipsets both support 20 PCI Express lanes: 16 lanes for the graphics card slots, 2 lanes for PCI Express x1 slots or integrated controller that use this interface and 2 lanes more for connecting to ATI SB450 or ULi M1573/1575 South Bridges. We should note that although the RADEON XPRESS 200 CrossFire Edition is positioned as an advanced solution for enthusiasts, the capabilities of the ATI SB450 are rather humble in comparison with the NVIDIA nForce4 SLI:
ATI RADEON XPRESS 200
North Bridge: ATI RD480
PCI Express x2 (1GB/s)
16bit / 1GHz
16bit / 1GHz
PCI Express lanes
Serial ATA ports
Parallel ATA channels
0, 1, 0+1
0, 1 (only for Serial ATA)
Secure Networking Engine
High Definition Audio (Azalia)
The only advantage of the ATI chipset is its support of the High-Definition Audio standard (Azalia) whereas the NVIDIA chipset uses an ordinary AC’97 codec. The SB450 is inferior to the NVIDIA chipset in the rest of the parameters as it supports only 8 USB 2.0 ports (against the nForce4’s 10), doesn’t have an integrated network controller and doesn’t support the Serial ATA-II standard. It means the manufacturers of RADEON XPRESS 200-based mainboards will have to use an external PCI Express or PCI network controller and integrated appropriate chips (for example, Silicon Image SiI3132) to fully support the newest Serial ATA-II hard disk drives. The nForce4 SLI would allow building a mainboard by adding only a PHY Ethernet controller.
A mainboard based on a RADEON XPRESS 200 CrossFire Edition chipset is needed to build a CrossFire platform and we are grateful to ATI Technologies for their giving us not only two RADEON X850 XT graphics cards but also an appropriate mainboard to perform our tests on. We traditionally use an AMD Athlon 64 4000+ processor in our testbed, so the mainboard we received was based on a RD480+SB450 chipset. The mainboard looks most impressive with its red slots and black lacquered PCB:
Ours was an engineering sample, with traces of colophony, numerous engineering connectors and jumpers and a not very convenient layout – the mainboard offers no expansion opportunities other than two PCI Express x16 and two PCI slots. The PCI slots are most inappropriately placed, too. The bottom PCI Express x16 slot is the Master, so the first PCI slot gets automatically blocked by the dual-slot cooling system of the RADEON X850 XT CrossFire Edition. It is the stranger then to see an abundance of free space under the second PCI slot – this part of the PCB is empty save for the diagnostic LED indicator with its Xilinx chip and an ATI logo! We think the PCI slots could be placed lower. It wouldn’t make the wiring of the PCB much more complex, but would make both slots available for use.
The PCI Express x16 are not configured on the software level as you might think looking at the empty space between them. The mode selector here is a special textolite terminator card which is installed into the top PCI Express x16 slot if there’s only one graphics card in the system.
If you do not use the selector, the bottom slot will work as PCI Express x8, causing a certain performance hit. Moreover, you have to activate the appropriate BIOS option after you’ve installed two graphics cards. We do not know if such a selector will be used in off-the-shelf mainboards. We think that at least some mainboard makers will offer software selection of the PCI Express modes since the selector is not convenient. It is not fixed in the slot and can be broken or lost, making it impossible to use the Master slot in the PCI Express x16 mode.
It is easy to explain why the mainboard lacks PCI Express x1 slots: two PCI Express lanes are used to connect the chipset’s Bridges, thus leaving only two lanes free. The South Bridge lacking an integrated network controller, one of the lanes is for the external 88E8052 Gigabit Ethernet controller from Marvell. The last lane is connected to a RAID controller SiI3132 which is responsible for the mainboard’s two Serial ATA-II ports. This controller is most appropriate here because as we found out during the installation of the operating system the four-channel Serial ATA controller integrated into the South Bridge is in fact two dual-channel SiI3112 controllers.
The developers of the SB450 must have thought it unnecessary to develop a Serial ATA controller from scratch but instead used a ready-made solution, integrating the appropriate logics into the chip. The development of the South Bridge was thus made cheaper and simpler, but stripped it of Serial ATA-II support (which may lead to compatibility problems with modern hard disk drives) and RAID 0+1 mode and also created a potential bottleneck since the SiI3112 uses a PCI interface rather than the faster PCI Express bus.
We didn’t perform any tests of the disk subsystem connected to the South Bridge of the RADEON XPRESS 200 CrossFire Edition since it would have been beyond the scope of this review, but we are sure it would be slower than a hard drive connected to the external controller.
A popular HDA codec ALC800 is employed in the audio section of this mainboard. The FireWire interface is also supported thanks to a VIA VT6306 chip. Thus, the reviewed mainboard is a high-class solution with support of all modern standards and interfaces, but leaving no room for further modernization.
The power connectors are placed properly. To improve stability in multi-GPU modes the mainboard has an additional Molex connector that reinforces the power circuit of the PCI Express x16 slots. ASUSTeK Computer uses an analogous solution in their mainboards – they call it EZ Plug.
As for the placement of the IDE and FDD slots, it is not perfect. The IDE slots are right behind the PCI Express slots and may become blocked by long graphics cards. For example, the IDE connector closest to the PCI Express x16 slots got blocked when we installed an NVIDIA GeForce 7800 GTX. The same problem may arise if ATI’s new-generation graphics cards turn to be as long as NVIDIA’s modern devices.
The FDD slot is located behind the 24-pin power connector. This is not very handy, but after all modern systems seldom use a 3.5” floppy drive. It should be acknowledged that wiring a multi-GPU-compliant mainboard is a daunting task as the engineers must put two PCI Express slots as well as many other components in a limited area. It is probably impossible to place each component conveniently for the user, especially with the classic “North Bridge + South Bridge” design, but approaching the ideal as closely as possible should be the goal.
The engineers took care about proper cooling of the mainboard’s components. The North Bridge is cooled by a tall needle-shaped heatsink with a small fan. Contrary to our apprehensions, the fan turned to be very quiet and it cools the not very complex RD480 chip – essentially a PCI Express controller – efficiently. Unfortunately, the North Bridge is placed too close to the CPU socket and makes it difficult to install a CPU cooler, especially if the latter uses the standard twisting clip. We managed to mount a Gigabyte 3D Rocket PCU22-SE cooler on the CPU, but with some serious difficulty. The South Bridge also has a small needle-shaped heatsink on (on thermal glue), while the power transistors of the CPU power circuit are grouped into a single block and are cooled with a rather big plate heatsink with a RADEON XPRESS logotype. The place usually occupied with the COM and LPT ports is empty. Off-the-shelf samples of the mainboard will probably have an exhaust fan there to take out the heat from the above-mentioned heatsink (the two-pin connector marked as PWR Fan indicates this possibility).
There are Power-On and Reset buttons below the Serial ATA-II controller, and the mainboard also has rubber feet which makes it an ideal test platform – you don’t have to search for a support and close the contacts with a screwdriver to start it up. :) The mainboard also carries a number of red LEDs that indicate the activity of each Serial ATA channel and report when the mainboard, CPU and memory modules receive power.
The current version of the mainboard’s BIOS (dated 08/12/05) is not free from some compatibility problems, Particularly, the mainboard refused to work with NVIDIA GeForce 7800 GTX graphics cards and with a 2GB OCZ Enhanced Latency DDR PC3200 Dual Channel Platinum memory kit – the system would only start up with one memory module installed or if we put two modules into the third and fourth DIMM slots, the memory would work as PC2100 only. So we had to use our time-tested 1GB OCZ Enhanced Bandwidth DDR PC-3200 Dual Channel Platinum memory kit which worked in our sample of the mainboard without problems.
Unfortunately, the mainboard was damaged during transportation. When we took it out of the package, we found that a corner of one of the IDE slots was split off and the empty bed (probably intended for a backup BIOS chip) was almost torn off its place. There was no other obvious damage, but the mainboard began to act up soon. For example, it refused to identify an Audigy 2 audio card. ATI replaced the damaged sample with a new one for us to finish our tests.
It’s too early to make any judgments about the mainboard. We had an engineering sample with a number of drawbacks described above. We hope that final versions of RADEON XPRESS 200 CrossFire Edition mainboards will come out without those drawbacks, or at least without the worst of them.
The RADEON X850 XT CrossFire Edition (left) doesn’t differ from ordinary RADEON X850 XT/XT Platinum Edition (right) externally, at least until you remove the cooling system.
One thing will catch your eye immediately: there is no S-Video connector and there is a DMS-59 connector instead of one DVI output. On a closer inspection you can notice a chip under the cooler which is missing on the ordinary RADEON X850 XT – the above-mentioned single-link TMDS receiver TFP401A from Texas Instruments. The remaining components of the frame-combining system are located under the cooler. Some differences can be seen on the reverse side of the PCB, too. There is no spot for a VIVO chip Rage Theater – there’s wiring pertaining to the Compositing Engine there, instead. There is also a 14-pin connector at the top of the PCB, probably for engineering purposes. It maybe serves for pre-programming the Xilinx matrix.
More points of difference rise up if you remove the cooler:
RADEON X850 XT CrossFire Edition
RADEON X850 XT/XT Platinum Edition
The left part of the PCB is all occupied by the Compositing Engine chips, the heart of which is the programmable matrix Xilinx Spartan-3. A 330MHz RAMDAC is located above it. To the right of the RAMDAC there is a TMDS transmitter that outputs the image on TFT panels with a DVI interface. The TV output controller integrated into the GPU is not used due to obvious reasons, and the developers didn’t want to burden the Compositing Engine with one more chip – it is already complex. So, you can’t connect your CrossFire to a TV-set with an RCA or S-Video input; you have to use an external coder if necessary. This is one more drawback of the current version of CrossFire technology.
The cooler on the card doesn’t differ from the one that first appeared on RADEON X850 series cards. It is the same turbine that sucks air from inside the system case and drives it through the copper heat-spreader that contacts with the GPU die and then exhausts the hot air to the outside. The solution is efficient, yet rather noisy. The plastic casing of the turbine works as a resonator, so the cooler remains quiet only at the lowest speed of the fan. There’s a wide gap between the heat-spreader’s sole and the case of the Xilinx matrix. This gap is filled with strange white foam-like thermal paste. Its efficiency as of a thermal interface is doubtful considering the thickness of the layer of the paste (about 1 millimeter). But maybe this is enough to cool the heart of the Compositing Engine.
This is where the difference between the RADEON X850 XT and the RADEON X850 XT CrossFire Edition ends. The right part of the PCB copies the design of the RADEON X850 XT to a smallest detail. We’ve got an R480 processor here with ordinary 1.6ns GDDR3 memory from Samsung. The clock rates of the two versions of RADEON X850 XT coincide and are 520MHz for the graphics core and 540 (1080) MHz for the memory chips.
As we have already said, the dual-slot reference cooling system from ATI operates really quietly only at the lowest fan rotation speeds. At higher rotation speeds the fan produces pretty loud noise. And since the cooling system casing is a good resonator, the noise acquires pretty unpleasant “plastic” shade, which turns out very irritating, especially during long term work with the system at this fan rotation speed. In regular work mode the cooler of RADEON X850 XT/XT Platinum Edition works at minimum fan rotation speed most of the time, so that the generated noise stays within reasonable range. It only speeds up during long work in 3D mode. CrossFire system acts differently: the master card fan rotated at very high speed throughout the entire test session, however, the worst thing about it was the constantly changing rotation speed, so that the noise produced by the system turned out simply unbearable. The only solution with a more unpleasant noise produced by the cooling system was XGI Volari Duo V8 Ultra. If you remember, we reported that its two high-speed fans generated high-frequency squealing noise (for details see our article called Club3D Volari Duo V8 Ultra Review: XGI Volari Family Coming to Graphics Market).
However, CrossFire cooling solutions does have one potentially great advantage: two graphics cards that suck the system air inside the PC case and then oust it outside the system should have a great positive effect on the internal system thermal conditions. I said “potentially great”, because you will hardly put up with the generated noise for the sake of better system cooling.
Of course, you should keep in mind that our test session was performed in an open test stand. However, if the mass CrossFire systems will act the same way as we have just described above, you will have to give up comfortable gaming experience once and for all.
Since it is right now impossible to overclock the CrossFire kit, we tested the overclocking potential of the graphics accelerators from this kit separately. Our tests revealed the following results:
The graphics cards worked stably at these frequencies. They received additional cooling from the 120-mm fan installed behind both cards. I have to say that 610MHz is a record for RADEON X850 XT in our labs. All our previous attempts to overclock the members of this graphics card family ended around 560MHz-580MHz.
We have also measured the power consumption of both graphics accelerators from the CrossFire system. Here we used a system based on Intel D925XCV Desktop Board, which was modified accordingly, so that we could measure the power on 12V and 3.3V lines of the PCI Express x16 slot. We also used a special adapter that allowed measuring the power send to the graphics accelerator via the 6-pin power connector. Our test system for this round of measurements looked as follows:
To create appropriate workload we ran FarCry 3D shooter Pier demo in 1600x1200 resolution with enabled FSAA and anisotropic filtering. When RADEON X850 XT CrossFire Edition is used as a single graphics card in the system, Compositing Engine doesn’t work. Therefore, the power consumption of this graphics accelerator remained the same as the power consumption of a regular RADEON X850 XT:
I doubt that the Compositing Engine requires more than 2-3W of power, so the difference in power consumption of the Master and Slave cards will anyway be negligible. Nevertheless, both cards working in a CrossFire system may consume up to 140W altogether. Of course, it is not so much compared with what two GeForce 6800 Ultra graphics cards would consume (155W), however it is a way higher than what a single GeForce 7800 GT would need (115W). ATI Technologies doesn’t recommend the use of any specific power supply units for CrossFire systems. However, keeping in mind our measurement results we would advise to go for high-quality PSUs with 400W+ capacity in continuous work mode. For example, a good choice in this case will be CoolerMaster RS-450-ACLY, Zalman ZM400B-APS or A.C. Ryan Ryanpower2 ACR-PS209. It would be best to avoid low-cost noname PSUs, because the specifications claimed on their stickers are very often different from the actual parameters. Besides, they are unreliable and do not guarantee stable output voltage because of too gravely simplified schematics.
The 2D image quality provided by the Slave card appeared exactly the same as that of any RADEON X850, i.e. pretty high for comfortable work in 1800x1440x75Hz. By the way, this graphics card features only one DVI-I ports: there is a standard 15-pin D-Sub connector instead of the second DVI-I port. I couldn’t figure out what pushed ATI to equip their slave card with this particular set of connectors: all analog monitors can be easily connected to the card via the corresponding adapter.
As for RADEON X850 XT CrossFire Edition, the situation appeared far not that simple here. The DVI-I output pot connected to directly the graphics processor (and the integrated RAMDAC) demonstrated the same image quality as that of the slave card. However, once we connected the DMS-59 via the CrossFire adapter, the maximum supported resolution appeared limited to 1920x1080 with the refresh rate of only 30Hz! For 1600x1200 the maximum supported refresh rate equaled 60Hz. It looks like the limitations imposed by the architectural peculiarities of the Compositing Engine influence the single graphics card mode as well, even though there is a RAMDAC with 330MHz frequency.
It doesn’t make much sense to evaluate the 2D image quality in CrossFire mode: the maximum supported resolution is only 1600x1200 with 60Hz refresh rate. And in lower resolutions the image quality simply cannot be bad. With multi-GPU mode enabled we noticed that the test screen we see before the OS boots up looked somewhat washed-out, as if we were running interpolation mode on a TFT display. It turned out that Compositing Engine does perform interpolation transforming the test VGA mode with 720x400 resolution into 1280x1024. The refresh rate in this case equals 85Hz.
CrossFire system was tested in the following test system:
For NVIDIA graphics cards we assembled a different test system. Here it is:
Following the standard testing procedure we adjusted the settings in NVIDIA’s and ATI’s drivers accordingly:
NVIDIA ForceWare 78.01:
We enabled full-screen antialiasing and anisotropic filtering from the game menu, if possible. Otherwise, we forced the necessary mode from the driver. We did not edit the games’ configuration files. We set the level of detail to the maximum in each game, so that it could be the same for ATI and NVIDIA solutions, but the rendering mode was selected depending on the capabilities of the given graphics card. If the game supported Shader Model 3.0, we used this mode for NVIDIA solutions. ATI cards worked in Shader Model 1.1/2.0/2.0b mode depending on the game.
We used the integrated benchmarks of those games that had them available. Of the results reported in the end of the benchmark showed not only the average but also the minimum values, we included them into the database, too. If the game had no integrated benchmarks we used FRAPS utility for our tests. In the latter case the minimum fps rates were also included into the database. Besides the RADEON X850 XT CrossFire kit, we have also tested the following graphics accelerators for a more in-depth and illustrative performance analysis:
Unfortunately, we didn’t have the chance to test an SLI configuration with two GeForce 7800 GT graphics cards, because there was only one graphics card like that at our disposal at that time. The tests were run in the following applications:
First Person 3D Shooters:
Third Person 3D Shooters:
CrossFire does provide certain benefits here, however, in pure mode it is not that significant at all, and doesn’t exceed 15-17%. Even in this case a single GeForce 7800 GT graphics card runs faster.
In eye candy mode, the situation is a little bit different: until 1600x1200 resolution CrossFire runs as fast as GeForce 6800 Ultra SLI, however, in 1600x1200 CrossFire suddenly dashes forward yielding only to GeForce 7800 GTX SLI. Compared with a single RADEON X850 graphics card we see a 35% performance improvement. It is not a dramatic increase, but among the SLI supporting platforms only pairs of two GeForce 7800 GT cards and two GeForce 7800 GTX cards appeared faster than CrossFire.
As the resolution increases, CrossFire gets more and more efficient. However, the best two RADEON X850 XT graphics cards can achieve is the performance level of a single GeForce 7800 GTX, and with FSAA disabled they can successfully compete with GeForce 7800 GT. The reasons for this are quite evident: we have a game using a lot of stencil shadows and GeForce 6/7 architectures supporting UltraShadow II. Nevertheless, I have to stress that the performance improvement provided by ATI CrossFire in Chronicles of Riddick is higher than in the previous game: 60%-70%.
As you probably know DoomIII is not very “ATI-friendly”, however, the use of new CrossFire technology allows reaching the heights of GeForce 7800 GT, and with enabled full-screen anti-aliasing CrossFire outperforms GeForce 7800 GTX and almost catches up with GeForce 6800 Ultra SLI. However, you should keep in mind that comfortable gameplay is possible only in 1280x1024 and lower resolutions, because you will be severely limited by the 60Hz refresh rate in 1600x1200, which is not very good for your eyesight, even though the performance will be quite good.
In d3dm4 level CrossFire proves even more efficient. In pure mode it is just a little bit behind GeForce 7800 GTX, and with enabled FSAA and anisotropic filtering it becomes the third fastest after the GeForce 7800 GTX and GeForce 6800 Ultra SLI systems.
If the game runs without full-screen anti-aliasing, CrossFire platform performs slower than a single RADEON X850 XT. You can feel the advantages of the multi-GPU mode only if you play at resolutions of 1024x768 and up with enabled FSAA. In this case RADEON X850 XT CrossFire is only 5fps behind GeForce 7800 GTX SLI.
In the Research level there is completely different gaming environment: the action mostly takes place indoors. RADEON X850 XT CrossFire is initially just a little bit slower than GeForce 7800 GTX, however once we switch to eye candy mode, it finally manages to get at the head of the race. The absence of Shader Model 3.0 support doesn’t allow ATI platform to defeat GeForce 6800 Ultra SLI, although it gets as close to the rival as 2fps difference in 1600x1200.
It turned out that CrossFire technology is not absolutely software-independent, and the results obtained in the F.E.A.R. demo version are a great example of that. The enabled multi-GPU mode slows down ATI CrossFire platform, although this is not a dramatic performance hit.
The situation in the Pariah game is truly amazing: while GeForce 6800 Ultra SLI gets close to 100% performance improvement, the CrossFire platform hardly runs 10% faster than a single RADEON X850 XT… Well, this is none other but a definite failure.
Almost in all work modes and resolutions the high-end graphics accelerators get held by the system CPU. Only in 1600x1200 with enabled FSAA and anisotropic filtering we can see some results: CrossFire platform falls a little bit behind a single GeForce 7800 GT: the performance gap makes about 3%-5%.
The same is true for the d3_c17_02 level. However, here we see that a team of two RADEON X850 XT graphics cards working in CrossFire mode gets about 2%-3% faster than GeForce 7800 GTX. As for the overall CrossFire efficiency, the performance improvement is not very high and doesn’t exceed 20%-22%. The maximum performance of 88fps cannot be achieved even in multi-processor mode.
CrossFire technology didn’t work in this game: the performance improvement was 0% even in the highest resolutions and with enabled anti-aliasing. Once again we can see that besides hardware issues, ATI CrossFire technology is not completely free from some software drawbacks.
The use of CrossFire support in this game resulted into a dramatic performance hit. We ran the tests a couple times to verify that this is not an accidental measuring error, however, the results remained the same. Why did CrossFire behave so strangely here, especially keeping in mind that they claimed its absolute software-independence? It might be the drivers that are to blame this time.
Just like in Half-Life 2, we see that high-end graphics cards easily reach their performance top in all resolutions. Nevertheless, in this one of the most popular games we can hardly see any performance improvement from the use of CrossFire technology in 1600x1200 with enabled antialiasing.
There is absolutely no performance increase in the Metallurgy level. Moreover, CrossFire system performs even slower than a single card would do. We notice some improvement in 1280x1024 in eye candy mode only. For a better comparison I would like to draw your attention to the fact that NVIDIA SLI works perfectly well here and demonstrates almost 70% performance improvement compared with a single-card mode.
Here we can see that both dual-graphics-card solutions demonstrate performance improvement compared with what a single graphics card would do. The improvement is not very significant for ATI CrossFire as well as for NVIDIA SLI: no higher than 17%-20% in 1600x1200.
In Splinter Cell: Chaos Theory , on the contrary, both technologies prove very efficient. In high resolutions with enabled full-screen antialiasing the efficiency of ATI CrossFire is approaching its theoretical maximum. And in general, the performance of the two RADEON X850 XT graphics cards with CrossFire technology is just a little lower than that of two GeForce 7800 GTX cards.
CrossFire also performs very well in rally auto simulator game – Colin McRae Rally 05 . Moreover, since the game is reach in relatively simple pixel shaders, which RADEON X850/X800 can handle just perfectly, the joint power of the CrossFire’s 32 pixel processors allows it to outperform even GeForce 7800 STX SLI in all modes and resolutions except 1600x1200 with enabled FSAA and anisotropic filtering.
The recent triumph was really impressive, and so is the current failure. All the way through LockOn we saw none other but a dramatic performance hit caused by enabled CrossFire support. It is remarkable however, that the minimal fps rate increased, i.e. the gaming performance improved noticeably, especially in higher resolutions.
Flight simulators seem to be far from CrossFire’s best. Nevertheless, we still managed to get pretty good performance improvement in higher resolutions, although in pure gaming mode only. The CrossFire platform fell about 15% behind GeForce 6800 Ultra in 1600x1200 resolution. If you are a dedicated flight simulator fan, I wouldn’t recommend going for CrossFire platform.
Pacific Fighters is the only game where we discovered visual artifacts in CrossFire mode. Every now and then the image would look like this:
For the first time we managed to get the performance as high as 40fps with maximum graphics quality settings in this very demanding strategic game. And this was only possible thanks to ATI CrossFire technology. Unlike the Pariah game, it is NVIDIA SLI that appeared useless here.
Two RADEON X850 XT in CrossFire platform allow us to hit the maximum performance in pure mode, and to catch up with GeForce 7800 GTX in eye candy mode. Overall, ATI CrossFire is more efficient in this game than NVIDIA SLI.
Multi-GPU solutions from ATI and NVIDIA act almost identically here in pure mode (that is with disabled FSAA). The systems with two RADEON X850 XT and two GeForce 6800 Ultra run almost neck and neck and demonstrate similar performance improvement compared with a single-card mode. With enabled FSAA CrossFire takes the leading position in higher resolutions. However, if we compare the performance improvement in this mode, then NVIDIA SLI will look more efficient than the multi-GPU technology from ATI.
You cannot benefit noticeably from CrossFire support in this game, so the laurels remain by GeForce 7800 GTX. However, to be fair, I should say that SLI technology is also of no particular use here.
Almost 20,000 points scored CrossFire – a great result keeping in mind that three tests of this benchmarking suite favor GeForce 6/7 based graphics solutions. Therefore, the gap of 844 points between our hero and NVIDIA GeForce 6800 Ultra SLI doesn’t look like a significant failure.
In Game 1 benchmark we observe certain parity between RADEON X850 CrossFire and GeForce 6800 Ultra SLI. NVIDIA solution is a little bit ahead in pure mode, while ATI takes the lead in eye candy mode with enabled full-screen antialiasing and anisotropic filtering.
GeForce 6800 Ultra SLI is the winner in Game 2 benchmark, because ATI CrossFire doesn’t have any special features improving its work with stencil shadows. And this type of shadows is used a lot in Game 3 and Game 3 tests of 3DMark03 package. In eye candy mode the gap between the leader and ATI CrossFire platform reaches 25% in 1600x1200 resolution.
We see the similar situation in Game 3 test, which only difference from the second one is higher vertex processor workload.
The only test in 3DMark03 suite where ATI RADEON X850 CrossFire manages to defeat GeForce 6800 Ultra SLI is Game 4 test. Our hero is an indisputable winner in pure mode and at least as fast as the rival in eye candy mode with enabled full-screen antialiasing. It is solely due to the peculiarities of ATI RADEON X850/X800 architecture, which is very efficient with relatively simple pixel shaders.
All in all, the results obtained in 3DMark03 are quite logical: the parity with GeForce 6800 Ultra SLI in Game1 and the impressive victory in Game 4 make up for the defeat in Game 2 and 3. Unfortunately for ATI, the NVIDIA GeForce 7800 GT SLI selling at about the same price point will be faster than RADEON X850 CrossFire.
As for the technology efficiency compared with the single-graphics card platforms, both multi-GPU technologies showed their best in 3DMark03, with just a slight advantage in NVIDIA’s favor. NVIDIA SLI is more software dependent, however it works to its utmost advantage if appropriate software optimizations take place, as its performance improves to the maximum in this case.
CrossFire platform performed very well in 3DMark05 . It fell only 1300 points behind GeForce 7800 GTX SLI, which is a very good result keeping in mind that NVIDIA had the total of 48 pixel pipelines against ATI’s 32.
3DMark05 test suite favors ATI’s solutions, which we can see immediately in the first test already: RADEON X850 XT CrossFire is as efficient as GeForce 6800 Ultra SLI in pure mode (no FSAA, no AF). As the workload increases, multi-GPU solution from ATI backs up and yields its positions to the rival. In this case it is nearly as fast as GeForce 7800 GTX, and is certainly supposedly slower than the GeForce 7800 GT SLI could be.
In Game 2 the vertex processors get loaded very heavily. ATI wins the race here, since CrossFire vertex processors work at a much higher frequency than the vertex processors of GeForce 6800 Ultra SLI: ATI CrossFire is the leader in all resolutions and work modes. The advantage is not so significant in eye candy mode as it is in pure gaming mode, but it is indisputable.
Game 3 looks very much like what we have just seen in Game 1 . At first RADEON X850 XT CrossFire outperforms GeForce 6800 Ultra SLI. However, when FSAA and anisotropic filtering are enabled the performance of ATI solution drops down to that of a single GeForce 7800 GTX. Since 3DMark05 runs by default at 1024x768 with disabled FSAA and AF, the total score is very logical: in this mode ATI RADEON X850 XT CrossFire wouldn’t yield to GeForce 6800 Ultra SLI, but would most likely fall behind GeForce 7800 GT SLI.
Before we draw our conclusions about ATI CrossFire technology, let us share our personal impressions of this platform with you first:
Just like its primary competitor, ATI managed to declare a lot of advantages for its CrossFire technology compared with the rivalry multi-GPU SLI. We criticized NVIDIA SLI technology and the developer’s praising promotion-like statements pretty severely back in the days. The things we said early this year remained acute up until now and we have to admit that they are also true for ATI CrossFire from many stand points.
The current incarnation of ATI CrossFire has one serious drawback for CRT monitor owners, namely, the maximum resolution it supports is 1600x1200 with 60Hz refresh rate. Unfortunately, it is also impossible to connect HDTV for CrossFire mode via the composite connector of RADEON X850 XT, which limits the application field for this technology even more. In fact, the only user group that will be able to use the 1600x1200 resolution on ATI CrossFire platform efficiently is the owners of large LCD panels.
RADEON X8 CrossFire Edition graphics cards can really be paired with any X800 and X850 cards from any vendors, which proves that this solution is highly flexible in terms of compatibility. However, if you have two graphics cards, one with 16 and another one with 12 pixel pipelines, we will actually get a pair of cards with 12 active pixel processors each and hence will not be able to support SuperTiling. As a result, there will be no tremendous performance increase from the second graphics accelerator in the system.
Although ATI introduced the so-called Compositing Engine, it didn’t completely eliminate the dependence of the system performance on the drivers in the multi-GPU mode. ATI CrossFire requires performance optimizations for each particular game, just like NVIDIA SLI. As a result, we can see the performance growing up high in some games, while remaining unchanged or even dropping down in other ones.
Just like with NVIDIA SLI 10 months ago, we cannot say that idea of CrossFire is bad or defective. However, the current situation is hardly favorable for getting a CrossFire based system. As for the solution offered by some ATI partners, namely to buy a second CrossFire Edition graphics card to upgrade the video system, it will hardly arouse a lot of enthusiasm, because many users will also have to replace their mainboard with a CrossFire one. It is much easier to go for a single GeForce 7800 GT, GeForce 7800 GTX or the upcoming RADEON X1800 XT, which is anticipated to be as fast as a pair of RADEON X850 XT cards but will not depend on the driver support for the given games and will definitely have no resolution limitations.
If we disregard the absence of any performance improvement in some games, there were no fatal compatibility issues with this platform. Once the mainboard damaged on the way to the lab was replaced, ATI CrossFire platform proved highly stable and reliable throughout the entire test session. All games would boot without any problems, except Pacific Fighters, and the image quality was identical to the reference. As for the performance improvement brought by the multi-GPU technology, ATI CrossFire cannot boast much unfortunately, especially against the background of more successful NVIDIA SLI:
Chronicles of Riddick
F.E.A.R. Multiplayer Demo
Painkiller: Battle Out of Hell
Unreal Tournament 2004
Prince of Persia: Warrior Within
Splinter Cell: Chaos Theory
Colin McRae Rally 2005
Lock On: Modern Air Combat
Warhammer 40.000: Dawn of War
Final Fantasy XI
+ performance improvement
- no performance improvement / performance hit
As you can see, ATI’s technology showed 0 or negative performance increase in 7 tests out of 21, while the competitor’s technology failed to cope with just 2 games and one semi-synthetic benchmark. Nothing to be surprised at, actually: SLI technology was introduced almost a year ago, so they definitely had enough time to bring it up right and make into a reliable solution. ATI Technologies also did have enough time to finalize their multi-GPU CrossFire technology: 4 months have passed since their preliminary announcement. It looks like 4 months turned out not quite enough. However, ATI has very little time left, as the R520 based product line is about to be released soon, too. And by the time the new graphics products are out, ATI CrossFire should be already at least as functional and efficient as NVIDIA SLI.
After a few postponed announcements, ATI RADEON X800/X850 CrossFire do not strike the High-End graphics accelerators market as a very powerful solution. Of course, ATI corrected the pricing of its X8 accelerators in order to meet the market realities, however, it will hardly make the new technology more attractive for the customers. According to the latest recommended pricing from ATI, CrossFire Edition graphics cards should be priced between $199 and $349:
This way the today’s top configuration is positioned to compete not really with the dual GeForce 6800 Ultra, but with mostly with a GeForce 7800 GT, which price has been dropped lately. And if in the first case ATI’s solution does look not bad at all, then in the second case RADEON X850 XT CrossFire will have really hard times, especially keeping in mind all those problems we have just talked about today.
As for the less expensive solutions, it would make sense to buy two graphics cards instead of one only in one case: if the two of them will cost no more than an analogous solution with one GPU onboard. Since ATI is going to roll out the whole new generation of graphics solutions in the nearest future and since NVIDIA may also exert some price pressure over the market, dual-card RADEON X800 based solutions will have very hard times competing with the single-chip champions. However, RADEON X800 XL CrossFire Edition and RADEON X800 CrossFire Edition will turn out so unpopular that none of ATI’s partners will be willing to ship them at all.
No matter if we like it or not, but multi-GPU is not just the spirit of the times, but a technology that is about to stay around for a long time. ATI followed NVIDIA on this one, other players, such as S3 Graphics are to follow next. Nevertheless, we should take into account that there are not that many users who really take advantage of multi-GPU technologies, so we shouldn’t overestimate their meaning.
ATI CrossFire today is NVIDIA SLI yesterday: driver and game issues will be identified and resolved with the time. In other words, it will take another couple of months before they manage to offer us a fully-fledged polished-off CrossFire driver.
However, the today’s CrossFire incarnation has one more serious drawback than just a raw driver: the resolution limitation of 1600x1200 at 60Hz refresh rate, which is in fact a hardware flaw. This is most likely to prevent ATI RADEON X8 CrossFire from getting really popular. Since the new ATI chips should have dual-link DVI transmitters, this limitation has every chance to disappear in the RADEON X1800 generation of products, leaving RADEON X8 and its drawback behind.
Another potential issue with ATI CrossFire may be the lack of CrossFire Edition graphics cards in the market. Moreover, this problem may touch not only upon RADEON X8 CrossFire Edition, but also upon the upcoming RADEON X1800 CrossFire Edition. Unless ATI decides to design its Compositing Engine as an add-on component and supply it with the corresponding mainboards, for instance.