by Alexey Stepin , Yaroslav Lyssenko
05/31/2009 | 11:32 AM
In a recent review we discussed the fastest dual-processor graphics solutions available for today’s games: ATI Radeon HD 4890 CrossFireX and Nvidia GeForce GTX 285 SLI. Besides making sure of their good performance in modern games, we also found out that ATI’s product was quite competitive to Nvidia’s one notwithstanding the difference in price. But for all their huge resources, these solutions are not the peak of the evolution of gaming 3D graphics hardware. Both GPU developers allow using more than two graphics processors in one subsystem. This opportunity has been available for a long time already. Nvidia’s products have offered it since GeForce 8800 GTX (G80) and ATI’s solutions have theoretically been capable of that since they acquired the CrossFire logic, i.e. since Radeon X1950 Pro (RV570). PC gamers are always interested in the latest developments in this field, so we are going to dedicate this review to today’s topmost multi-GPU configurations.
At this moment, AMD’s and Nvidia’s products can work in graphics subsystems with up to 4 graphics cores, but the maximum configuration in either case can only be assembled out of a pair of dual-processor cards: ATI Radeon HD 4870 X2 or GeForce GTX 295. There is one important note, however. The former card is inferior to a tandem of two Radeon HD 4890 cards that have higher GPU and memory clock rates whereas the latter has a cut-down configuration in comparison with Nvidia’s single-chip flagship GeForce GTX 285. Therefore 3-way configurations seem to be more optimal at the moment and we will focus on them in this test session (moreover, we don’t have two samples of GeForce GTX 295).
We will benchmark 4-way configurations as soon as Nvidia has released an updated GeForce GTX 295 with a single-PCB design and AMD’s partners have introduced Radeon HD 4890 X2, but in this review we will compare AMD’s 3-way CrossFireX and Nvidia’s 3-way SLI platforms. We will check out their performance in modern games, noisiness, power consumption and stability. We will also discuss the long-debated problem of the quality of transparent textures antialiasing. Let’s get started.
Noisiness is an important characteristic of a modern graphics card. No one wants a graphics card that roars like a taking-off plane even if it beats every other card in sheer speed. This is especially important for super-fast multi-GPU subsystems, so we measured the amount of noise they produce.
The results are discouraging. The ambient noise level equaling 43dBA, the Nvidia GeForce GTX 285 3-way SLI configuration is as loud as 60dBA in 3D applications. This is far from comfortable. The Radeon HD 4870 X2 4-way CrossFireX is not much better. Although it has two rather than three fans, the cooling system of the Radeon HD 4870 X2 has been criticized for producing too much noise. The 3-way configuration based on Radeon HD 4890 is the quietest of all, yet not quite comfortable, either.
These results come as no surprise because graphics cards in top-speed multi-GPU subsystems reside close to each other and work in very cramped conditions where they can get but little fresh air for cooling. This is especially true for the GeForce GTX 285 3-way SLI. The system having no air for breathing, the GPUs and other components quickly get hot and the cards have to step up the speed of their fans to avoid overheat. And even this may sometimes be useless. Our testbed would occasionally overheat and hang up during our tests and we had to wait for it to cool down to go on with the benchmarks.
This proves that the development of gaming 3D graphics hardware has almost reached a dead-end. Even installing as many as three top-class expensive graphics cards into your computer cannot ensure an acceptable speed in games like Crysis but brings about stability and compatibility issues and a huge increase in noise, power consumption and heat dissipation. Anyway, if you dare assemble one of the abovementioned systems, you must make sure beforehand that your system case can accommodate three dual-slot graphics cards and has good ventilation. Otherwise, you’ll have a lot of problems.
Our power consumption testbed cannot measure the electric parameters of more than one graphics card. Therefore, the numbers below are estimates calculated by summing up the power consumption of the individual cards in 3D mode.
The real numbers are going to be somewhat different because the power consumption of the Slave card is not zero even in 2D mode, but you can get a general notion of how “economical” top-performance multi-GPU solutions are.
The numbers agree with our opinion that multi-GPU technologies are a dead-end in the evolution of gaming graphics hardware. The Radeon HD 4890 3-way CrossFireX needs as much as 350 watts under load whereas the Radeon HD 4870 X2 tandem needs more than 500 watts. And they do not even guarantee a comfortable frame rate if the game developer has focused on modern visual effects but forgotten about code optimization. One thing is clear, though. If you are still bent on building a multi-GPU graphics subsystem, you should get a high-quality power supply with wattage rating of 750W and higher.
We are going to investigate the performance of multi-GPU graphics systems in contemporary games using the following testbed:
The graphics card drivers were configured in the following way:
Nvidia GeForce:
ATI Catalyst:
The list of benchmarks includes the following gaming titles and synthetic tests:
First-Person 3D Shooters
Third-Person 3D Shooters
RPG
Simulators
Strategies
Semi-synthetic Benchmarks
We selected the highest possible level of detail in each game using standard tools provided by the game itself from the gaming menu. The games configuration files weren’t modified in any way, because the ordinary user doesn’t have to know how to do it. We made a few exceptions for selected games if that was necessary. We are going to specifically dwell on each exception like that later on in our article.
The following graphics accelerators and multi-GPU configurations participated in our today’s test session:
We ran our tests in the following resolutions: 1280x1024, 1680x1050, 1920x1200 and 2560x1600. Everywhere, where it was possible we added MSAA 4x antialiasing to the standard anisotropic filtering 16x. We enabled antialiasing from the game’s menu. If this was not possible, we forced them using the appropriate driver settings of ATI Catalyst and Nvidia GeForce drivers.
Performance was measured with the games’ own tools and the original demos were recorded if possible. We measured not only the average speed, but also the minimum speed of the cards where possible. Otherwise, the performance was measured manually with Fraps utility version 2.9.8. In the latter case we ran the test three times and took the average of the three for the performance charts.
To cater to our readers’ requests, we performed an express test of the quality of antialiasing. An additional reason for us to do it was the introduction of 32x FSAA in the new drivers from ATI and Nvidia.
The drivers were configured as follows:
As usual, we captured the screenshots in Fallout 3 and Half-Life: Episode Two because these games have no problems with turning antialiasing on and abound in such graphics as fences, hanging wires and other micro-geometry that help evaluate the quality of FSAA.
Let’s first check out the new 32x FSAA mode in its ATI’s implementation:
|
|
The difference is insignificant although the 32x mode does provide a somewhat higher quality of antialiasing of small details such as ropes, struts and barbed wire thanks to the larger amount of color samples. These improvements are inconspicuous when you are actually playing the game unless you are looking for them deliberately.
|
|
The Episode Two screenshots speak for themselves, yet the image quality improvements are not as large as to justify such resource-consuming FSAA algorithms. The effect is really negligible if you compare it to the effect from switching from no FSAA to 4x FSAA.
Now let’s see how Nvidia’s solutions behave:
|
|
The difference is even less conspicuous in Fallout 3 than with ATI’s products and is definitely not worth the performance hit provoked by such resource-consuming antialiasing methods.
|
|
Like in Fallout 3, the effect from FSAA 32xQ is not worth the performance hit. This serves to confirm our point that 4x FSAA is the satisfactory level of antialiasing for most users and applications. The most fastidious gamers can try 8x FSAA (8xQ FSAA in Nvidia’s terminology) that ensures a nearly ideal quality of antialiasing at an acceptable performance hit, especially on top-class graphics cards.
Besides 32x FSAA, we will compare the quality of antialiasing of transparent textures in ATI and Nvidia solutions’ Adaptive AA and Transparency AA modes, respectively. This will help us find the ultimate answer what modes to use for our tests of graphics cards’ performance. Like in the previous case, ATI’s products go first.
| |
|
|
It is virtually impossible to find a difference between the lack of adaptive antialiasing and the Adaptive AA:Performance/Quality modes even if you come up to the mesh fence. When we compared the Performance and Quality modes directly in The Compressonator, the resulting screenshot (produced by deducting the original screenshots from each other) was a black screen, meaning that the originals were 100% identical. But perhaps it is only Fallout 3. Let’s see what we have in Half-Life 2: Episode Two.
| |
|
|
There are differences here, even though negligible ones. Take a look at the spruces’ branches: Adaptive AA improves their rendering quality even in Performance mode, although does not make it ideal. When we compared the spruces in Adaptive AA: Performance and Quality modes, we again got a black screen as the result. Thus, Adaptive AA is itself beneficial but there is absolutely no difference between the different Adaptive AA modes: Performance and Quality provide a same-looking picture.
Now it’s Nvidia’s turn:
| |
|
|
As opposed to ATI’s solutions, the difference is far more obvious: when you turn MS TAA on, the fence is displayed more neatly, with fewer undesired gaps. The fence looks sharper than on the ATI Radeon HD. It improves even more in the SS TAA mode and becomes better-looking than in the Adaptive AA modes on the ATI Radeon HD cards.
| |
|
|
Like with the ATI Radeon HD solutions, the effect from the enabled Transparency AA is conspicuous in the spruces in the first place. The antialiasing quality provided by MS TAA is roughly equal to ATI’s Adaptive AA. The quality of SS TAA is somewhat higher but not as conspicuously higher as in Fallout 3.
Thus, Nvidia’s Transparency Multisampling is roughly equal to ATI’s Adaptive AA in terms of quality. Therefore our choice of transparency antialiasing settings for our gaming tests is justifiable as it ensures equal visual quality for all products. We don’t see any reason to change our method as yet.
And now we can proceed to our gaming tests.
Starting from version 1.3 we use the game’s integrated benchmarking options together with a custom demo record. Unfortunately, this method does not report the bottom frame rate.
The multi-GPU systems included into this test session are premium-class solutions and should be compared at display resolutions of 1920x1200 and higher. When the resolution is lower, these platforms are obviously redundant.
The Radeon HD 4870 X2 tandem wins the popular gaming resolutions but produces so much noise and consumes so much power that we would prefer the Radeon HD 4890 3-way CrossFireX for practical purposes. As for single-PCB dual-core products, the GeForce GTX 295 is the overall leader as it is quieter and somewhat faster than the Radeon HD 4870 X2.
It is different in games running on CryEngine 2: you need the most advanced and expensive cards and multi-GPU configurations in order to get a playable speed in them even at 1680x1050. If you want to enjoy the highest graphics quality settings, that is.
The single-PCB products are not quite good at that task. The GeForce GTX 295 stops short of the desired 25fps and you need a GeForce GTX 285 SLI tandem or an appropriate 3-way SLI configuration in order to keep the frame rate above that level. The 3-way SLI platform ensures a larger reserve of speed but also produces much more noise. We don’t think it is a good solution especially as it cannot maintain an acceptable bottom speed at 1920x1200, either.
As for ATI’s products, their playability is limited to 1280x1024. Note that the Radeon HD 4890 3-way CrossFireX subsystem has a low bottom speed. It must be having problems synchronizing its three GPUs.
We disabled the integrated frame rate limiter in the game console for the sake of comparing the cards. The game’s built-in benchmarking options do not provide information about the bottom speed, so there is no such info in the diagrams.
Using multi-GPU configurations for Quake Wars does not make sense because even the ordinary Radeon HD 4890 ensures a twofold reserve of speed at 2560x1600. To remind you once again, the game’s network engine synchronizes all events with a frequency of 30Hz, so the maximum frame rate is limited to 30fps.
Anyway, we can note that starting from the resolution of 1680x1050 this test is won by the GeForce GTX 3-way SLI system. It enjoys a 22% lead over the second-best Radeon HD 4870 X2 4-way CrossFireX at 2560x1600.
Despite the obvious kinship, Far Cry 2 is a far less demanding application than Crysis Warhead, so every tested solution copes with the resolution of 2560x1600. The only exception is the single Radeon HD 4890 which must have been limited by its 16 RBEs.
The GeForce GTX 295 is the overall winner if you consider the whole of its consumer properties whereas the GeForce GTX 285 SLI looks like an optimal compromise between highest performance and noise. The addition of a third card to the SLI platform increases the average frame rate by 24% at 2560x1600, but we don’t think that this performance growth is worth the louder noise, especially as it is next to impossible to perceive that growth in the process of gaming.
The GeForce GTX 285 3-way SLI and Radeon HD 4870 X2 4-way CrossFireX are in the lead among the premium-class solutions. The Radeon HD 4890 3-way CrossFireX is somewhat slower but remains our favorite in the heavyweight category due to its quieter operation. The discrete dual-processor Radeon HD 4890 CrossFireX and GeForce GTX 285 SLI platforms go neck and neck, so we don’t see a clear winner in this segment. The GeForce GTX 295 expectedly wins among the integrated dual-chip cards.
Anyway, the ordinary single-processor GeForce GTX 285 is going to be the best choice for playing the game at 2560x1600 without such problems as incompatibility or much noise. This card is quieter and more economical even in comparison with the GeForce GTX 295, let alone the more advanced solutions. And it does not depend on the software support for multi-GPU configurations.
The game runs on the Source engine and has an integrated benchmark, but the latter does not report the bottom speed information.
The multi-GPU platforms are redundant for Left 4 Dead. A single Radeon HD 4890 is quite enough for playing the game comfortably at 2560x1600 with maximum graphics quality settings together with 4x FSAA. You can try using more resource-consuming antialiasing modes, but the resulting image quality improvements are often so negligible that you have to look for them specifically. Gazing at fences to see any difference in image quality may be dangerous during actual gaming.
Still, we can note that the Radeon HD 4870 X2 4-way CrossFireX subsystem is ahead of the GeForce GTX 285 3-way SLI here. The GeForce GTX 295 is the best choice among the single dual-processor cards, so ATI’s supporters have to wait for any products like Radeon HD 4890 X2. We hope such graphics cards will come out soon and be equipped with a better cooler than the Radeon HD 4870 X2.
To achieve a playable speed in this game we disabled FSAA and such resource-consuming options as Sun rays, Wet surfaces and Volumetric Smoke. We use the Enhanced full dynamic lighting (DX10) mode for our test and additionally enable the DirectX 10.1 mode for the ATI cards.
Notwithstanding the considerably relaxed settings, not all of the tested solutions can cope with the game at high resolutions. Both single-processor cards are limited to 1680x1050 only.
Multi-GPU solutions are usually bought to play at the highest resolutions, but it is hard to see which is better among them here. Each of the three subsystems with two and more GPUs delivers the same speed, but the Radeon HD 4890 3-way CrossFireX is preferable in terms of noisiness. Moreover, the latter boasts the best result at 2560x1600, even though the gap from its opponents is very small.
Like in many other tests, top-performance multi-GPU solutions deliver similar results at every resolution from 1280x1024 and higher. And our opinion has not changed: if you are interested in assembling a top-speed gaming platform with more than two GPUs, you should consider the configuration with three Radeon HD 4890 in the first place as it is the quietest and most economical among the platforms that equal each other in terms of performance. On the other hand, even a single Radeon HD 4890 is quite enough even for playing the game on a 30-inch monitor.
Prince of Persia is a more demanding application and the single Radeon HD 4890 cannot deliver a playable speed at 1920x1200. The GeForce GTX 295 can’t cope with the task at 2560x1600 whereas the GeForce GTX 285 SLI and Radeon HD 4890 CrossFireX tandems are sufficiently fast then and have rather acceptable noise and power consumption parameters.
Yes, the more advanced configurations ensure a large reserve of speed but the Radeon HD 4890 3-way CrossFireX is the only one of them which is more or less comfortable for your ears. And that configuration is somewhat inferior to the Radeon HD 4870 X2 tandem as well as to the GeForce GTX 285 3-way SLI.
Advanced multi-GPU solutions are so redundant for Fallout 3 and its add-ons that we cannot see a clear winner even at 2560x1600. All of them deliver the same performance. The GeForce GTX 285 SLI offers the best bottom speed but the Radeon HD 4890 CrossFireX outperforms it in terms of average frame rate. As you can see, there is almost no effect from adding a third or fourth GPU into the graphics subsystem: when switching from two to three GPUs the performance growth is a mere 2% for the Radeon HD 4890 3-way CrossFireX whereas the GeForce GTX 285 3-way SLI is even 3% slower than its 2-way counterpart.
Mass Effect becomes a highly demanding application if you force 4x FSAA on. It is only with three or more GPUs that you can get a playable frame rate at 2560x1600. The GeForce GTX 285 3-way SLI is in the lead, being 17% ahead of the Radeon HD 4890 3-way CrossFireX. ATI’s solution is quieter and more economical, though.
We have to say it once again: even dual-processor solutions are overkill for this game at every display mode including 2560x1600. Still we can note that the Radeon HD 4890 3-way CrossFire is in the lead and is also the quietest and most economical in its class (if words like quiet and economical can be applied to such solutions at all).
We use the in-game benchmarking tools that do not allow to measure the bottom frame rate. We also enable DirectX 10.1 support for ATI’s solutions.
The multi-GPU solutions show their potential and are indeed useful in this game if you want to play at resolutions of 1920x1200 and 2560x1600. The multi-processor products vary in scalability, though. The Radeon HD 4890 platform has a 15% performance growth when switching from two to three GPUs at 2560x1600, but the GeForce GTX 285 platform offers a twofold performance increase then! When a fourth GPU is added to the RV770-based platform, its performance increases by 60% more.
Unfortunately, we could not benchmark the GeForce GTX 295 pair in Quad SLI mode but such a subsystem would probably be limited by the amount of local graphics memory at 2560x1600 as is indicated by the results of the single GeForce GTX 295.
The game has a frame rate limiter fixed at 30fps.
Nvidia’s solutions have problems in Red Alert 3 when you turn on full-screen antialiasing but you don’t have to worry if you’ve got three GeForce GTX 285 cards: they are 15% ahead of the single Radeon HD 4890 at 2560x1600. The only inconvenience is that the 3-way SLI subsystem is going to roar like hell.
Of course, this is an example of poor code optimization because the GeForce GTX 285 surely has all the computing and texturing resources necessary to run this game fast. The Radeon HD 4890 can’t be that much better than its opponent.
The recently released add-on to the original game does not introduce any technical innovations but contains a new plotline that allows you to play for the USSR.
3- and 4-GPU solutions may be interesting for people who want to play World in Conflict at 2560x1600 as they can provide an acceptable bottom speed then. The GeForce GTX 285 3-way SLI is the leader among such platforms. It is 34% ahead of the ATI Radeon HD 4890 3-way CrossFireX. Although behind the leader, ATI’s solution delivers a playable frame rate, too. And it is better in terms of noisiness, power consumption and heat dissipation.
The GeForce GTX 285 3-way SLI is not very good in 3DMark06. It can’t score even 20,000 points. The Radeon HD 4870 X2 scores 22,000 points while the Radeon HD 4890 3-way CrossFireX stops just a little short of that mark.
The individual tests produce more interesting results: ATI’s solutions seem to hit the performance ceiling imposed by the testbed’s CPU in the SM2.0 tests. The GeForce GTX 285 3-way SLI is about as fast as the GeForce GTX 285 SLI pair and is even inferior to the single GeForce GTX 295. In the SM3.0/HDR tests the graphics subsystems are ranked according to the number of GPUs, but we can see that Nvidia’s systems do not accelerate when they switch from two to three GPUs. With ATI’s platforms, adding a third GPU leads to a performance increase but adding a fourth GPU gives you nothing.
We minimize the CPU’s influence by using the Extreme profile (1920x1200, 4x FSAA and anisotropic filtering). We also publish the results of the individual tests across all display resolutions to provide a full picture.
As opposed to 3DMark06, 3DMark Vantage produces logical results: the more GPUs in the graphics subsystem, the faster it is. The GeForce GTX 285 3-way SLI is an expected winner because it supports PhysX acceleration whereas ATI’s solutions do not. Take note that the platform with three Radeon HD 4890 cards is not much inferior to the system with two Radeon HD 4870 X2, which is an indication that the RV770 is going to be soon ousted by the RV790 in the same way as the G200 has been replaced by the more advanced G200b.
In both tests the GeForce GTX 285 3-way SLI platform comes out the winner, just like in the overall scores. ATI’s solutions are slower due to the lack of PhysX support, but the Radeon HD 4890 3-way CrossFireX is comparable to the 4-way configuration built out of two Radeon HD 4870 X2.
Recently we benchmarked dual-processor tandems Radeon HD 4890 CrossFireX and GeForce GTX 285 SLI and referred to them as heavyweights. So, this review has been about super heavyweights that feature the combined power of three or even four GPUs. But do such systems make a practical sense or they should only be used for setting speed records in benchmarks? We will sum up this test session and answer the question now.
It is clear that modern multi-GPU solutions are redundant for the resolution of 1280x1024 (they are used for that resolution but rarely, though). There are but few exceptions, particularly the CryEngine 2 based games such as Crysis and Crysis Warhead. In these heavy applications the GeForce GTX 285 3-way SLI subsystem can achieve an average frame rate of 55-60fps and a bottom speed of 29-30fps. Besides that, this advanced and expensive subsystem can be useful for the highly demanding flight sim Tom Clancy’s H.A.W.X. – if you don’t mind the noise it produces.
The resolution of 1680x1050 pixels is somewhat more interesting although we don’t think that people who have inexpensive 20-22-inch monitors will shell out $700-1000 for the graphics subsystem. Crysis Warhead, H.A.W.X., and Mass Effect with forced 4x FSAA are the only games that can utilize the advantages of the 3- and 4-way graphics subsystems. In every other case a single Radeon HD 4890 or GeForce GTX 285 is going to be enough.
The display resolution of 1920x1200 pixels is still rather rare even though there are inexpensive 24-inch TN-based monitors, e.g. Samsung SyncMaster 245B or Acer AL2416W. Users who have 1920x1200 monitors are more likely to be able to afford two or more top-class graphics cards.
Ironically, the 3-way GeForce GTX 285 configuration cannot maintain an acceptable bottom speed at Crysis Warhead. The discrete multi-GPU solutions show their benefits in S.T.A.L.K.E.R.: Clear Sky instead. They ensure a larger reserve of speed than the single GeForce GTX 295. It is hard to name the best solution here but we’d prefer the Radeon HD 4890 3-way CrossFireX due to its better noise characteristics. The same goes for Mass Effect and H.A.W.X.
The resolution of 2560x1600 is the paradise for many devoted gamers because a 30-inch monitor allows for a deeper immersion into the virtual world. It is here that super-fast multi-GPU solutions show their full potential.
The GeForce GTX 285 3-way SLI is not always in the lead at 2560x1600. It wins 9 and loses 6 tests to the Radeon HD 4870 X2 tandem but it is only in four tests (Crysis Warhead, Enemy Territory: Quake Wars, Mass Effect and World in Conflict: Soviet Assault) that the gap is over 15%. Nvidia’s platform has a large advantage (the score is 12 to 3) over the Radeon HD 4890 3-way CrossFireX but the gap is large in five tests only.
Is it so bad for ATI’s solutions, anyway? The difference is insignificant in Crysis Warhead because Nvidia’s 3-way platform cannot deliver acceptable speed, either. In the other tests the Radeon HD 4890 3-way CrossFireX provides a playable frame rate even though being slower than its opponent. Moreover, the Radeon HD 4890 3-way CrossFireX is far more comfortable in terms of noisiness. Delivering high performance at 2560x1600 at a rather comfortable noise and the lowest power consumption in this class, the Radeon HD 4890 3-way CrossFireX looks like the most optimal solution to us. The Radeon HD 4870 X2 4-way CrossFireX is worse because its advantage over the 3-way subsystem based on newer RV790 cores is small whereas its power consumption and noise level are higher.
Winding this review up, we want to say it once again that today’s multi-GPU technologies have a number of fundamental drawbacks that cannot be corrected even theoretically. Modern multi-GPU solutions depend on software support from both game developers and driver programmers. The so-called heterogeneous multi-GPU might save the day, but this concept seems to have been abandoned without making it into hardware implementations. Perhaps DirectX 11 will change something, but it’s yet too early to make any guesses about it. And beside everything else, multi-GPU solutions are clumsy, noisy, power-hungry, and very expensive.
We do think that a buyer who spends over $500 for one graphics card or multiple cards should get maximum performance out of the box, without bothering about versions and settings of the drivers and the availability of SLI/CrossFireX optimizations for his favorite games. Game developers who cannot make good use of the huge resources of modern graphics hardware are to blame, too. The only positive trend is that there are more and more multiplatform projects that have to be optimized for the modest resources of today’s gaming consoles.
