Tech-Hounds.com

Crossfire: ATI's Answer to SLI - Part 1

It's been a long time coming, but we finally decided to give in and test a Crossfire setup. For those of you out of the loop, Crossfire is ATI's take on the multi rendering solution, very much like NVIDIA's SLI. We were a little bit skeptical of multi rendering solutions in general, more so after looking at what SLI has to offer (you can read our article on SLI here and here). Add to the fact, reviews and previews of Crossfire first appearance in ATI's X800 series was quite disappointing to say the least.

Much have happen since that day. Today's Crossfire implementation can be considered 'version 2.0' of Crossfire. There are several obvious difference, one of which is the use of a new motherboard chipset (from the Radeon Xpress 200  to Radeon Xpress 3200). However, the most important factor, at least for mainstream parts (the Radeon X1600 series) is that you don't need a master card anymore. The first version of Crossfire require users to purchase a special master card in addition to an ordinary Radeon X800 / X850 series. So now users can pick any Radeon X1600 series card, pair them up in a Crossfire motherboard and enjoy what Crossfire has to offer. Also gone is the Crossfire cable / dongle. Now, all card to card communications are done through the PCI Express bus. To alleviate concerns of bus congestion, ATI have armed the new chipset with 2 full speed x16 PCI Express slots. With these changes, ATI has addressed much of the gripes of many reviewers and users concerning Crossfire.

However, for high end cards such as the Radeon X1800 / X1900 series, you still need a master card and the Crossfire cable / dongle (although you apparently don't need a master card for X1800GTO). The reasons for this is unclear.  So it looks like, users who want a dongleless high end Crossfire solution from ATI will likely have to wait for the next generation of graphics cards.

Since reviews and technological analysis of Crossfire are already abundant on the Internet, we won't go into the technical details of Crossfire. For a detailed analysis, you could read Beyond3D's preview here. We will just touch this subject very briefly.

Although it has undergo some changes, the heart of Crossfire is relatively unchanged. The basic premise is very much the same as NVIDIA's SLI - to split the rendering load (equally) across two cards. This is usually done by splitting the screen, with one card rendering one part and the other handling the second part. Keep in mind, when we're talking about the screen, we don't mean the area, rather the load. That's because the lower parts of the screen tend to have more objects than the upper part. Here Crossfire differs from SLI. In addition to scissor mode (splitting the screen into two parts by load) and alternate frame rendering (the two card alternatingly renders the screen), Crossfire supports another load balancing mode - supertiling. However, most of the performance gain should come from the use of alternate frame rendering mode.

For more quality oriented users, Crossfire offers Super AA - similar to SLI AA mode. Super AA also works much in the same way. Both cards will render the same frame with AA, but with slightly different sample patterns. ATI is quick to point out that due to the programmable sampling hardware present in their hardware, Super AA sample patterns are not just two similar sample pattern with slightly different placements.. Theoretically, ATI's Super AA 8x mode (each card using 4x AA) is different and have better image quality in terms of less aliasing than NVIDIA's SLI 8x AA. In addition to the obvious AA modes (Super AA 8x and 12x), there are two other modes that might be very interesting to look at - 10x and 14x. Super AA 10x is actually computed from samples taken from Super 8x AA sample pattern and 2 additional samples, but these 2 additonal samples are computed with super sampling. For the less technical savvy users, that means what it offers is pretty much similar to NVIDIA's 8xS (roughly). In you case you're wondering, Super AA 14x is Super AA 12x extended to account for 2 samples SSAA.

We'll be looking at what these Super AA modes has to offer in the second part of this article. For now, we will be focusing on performance. Just like our SLI article, we've decided to start from the low end of the spectrum - a pair of Radeon X1600XTs. Granted, most gamers out there considering Crossfire or SLI will likely use the much faster, high end cards such as Radeon X1900XT or GeForce 7900GTX. However, by choosing a less powerful graphics card(s), we can see what Crossfire has to offer in terms of performance but on a much less system limited setup. So, if Crossfire does work, performance can still go up without us hitting the limits of our processor. This way, we can focus this article on the performance aspects of Crossfire.

Performance

AA and AF settings are applied in game, whenever possible. We had to apply AA and AF from the driver panel for Call of Duty, Homeworld 2 and Richard Burns Rally. In Quake 4, we enable both AA and AF from the console, with the command 'r_multisample' set to 4 and 'image_anisotropy' set to 16. V sync was disabled both from inside the game and on the driver panel.

Call of Duty, Homeworld 2 and Richard Burns Rally was configured to the highest possible detail. F.E.A.R settings are pretty similar to our usual test settings (all settings set to maximum / on except for soft shadows), but for today's test we opted to set the volumetric lights to 'Off'. After all, we're only testing with two Radeon X1600XTs. Quake was set to 'High Quality' while Serious Sam II was tested with our usual standard settings (maximum settings but with custom resolution, anti aliasing and anisotropic filtering settings). The same also applies to Splinter Cell Chaos Theory - the game was benchmarked to use its shader model 3.0 and everything turned on, except for HDR which is only turned on for HDR performance tests.

We ran our usual tests under four different configurations - both in single and Crossfire mode and with and without Catalyst AI enabled (set to 'Standard'). We had to do this since Crossfire will work best when Catalyst AI is enabled. When enabled, Catalyst AI will apply the best Crossfire mode (alternate frame rendering) for a particular game if possible. If not, Crossfire fall back to its default rendering mode (either supertiling or scissor mode). However, Catalyst AI also perform other optimizations, so it's better to only compare results with the same setting (Catalyst AI enabled). Results with Catalyst AI disabled are included for information only.

We'd like to thank ASUS for supplying the ASUS A8R32-MVP Deluxe motherboard, Gigabyte for supplying the two Radeon X1600XT Silent Pipe II and Tagan for supplying the additional power supply for this article.

Our test setup
AMD Athlon 64 3500+ socket 939
2 x 1024 MB Kingston KVR 3-3-3 PC3200 DDR-SDRAM
ASUS A8R32-MVP Deluxe Radeon Xpress 3200 Crossfire motherboard
2 x Gigabyte Radeon X1600XT Silent Pipe II DDR3 256 MB graphics card (GV-RX16T256V-RH)
Maxtor DiamondMaxPlus9 80 GB Serial ATA 8 MB buffer
ASUS E-616 DVD-ROM
Tagan TG530-U15 530 watts ATX/BTX power supply

Windows XP Professional with Service Pack 2 installed
ATI Catalyst 6.5 reference driver
Realtek High Definition Audio 5.10.00.5202 driver.
DirectX 9.0c

Performance

The results:

The graphs are pretty self explanatory, but in case you can't see the text, they're arranged in groups of three colors: green for minimum fps, blue for average fps and red for maximum fps. The first group represent test results under default settings, and the second represent results with 4xAA and 16xAF enabled.  For HDR tests, the second groups represent results with HDR and 16x AF. These results are in frame rate per second.


We can see Crossfire is really paying dividends here, even if you're only using a pair of Radeon X1600XTs. The difference in average fps is not that much (5 fps without AA and AF), but overall the Crossfire setup is faster. With AA and AF, the difference grew to around 30 fps or about 20 percent faster than a single card. At higher resolutions, the difference grew to about 25 percent at 1280 x 1024 and 60 percent at 1600 x 1200, with AA and AF. However, even in Crossfire mode, a pair of Radeon X1600XTs can still run out of gas - it can't really maintain a high minimum fps above 1024 x 768.

Call of Duty is obviously one game that has a profile embedded in the drivers. This lets the Catalyst drivers select the best rendering mode (AFR) with Crossfire, however we can also see there's some other optimizations at work here. Notice that the differences we're talking in the paragraph above about are from single and multi configuration with Catalyst AI enabled. Those improvements are 'pure' Crossfire related. If you look at the differences between single configurations results with and without Catalyst AI enabled, Catalyst AI optimizations gave us about 5 - 10 fps more in default settings and 10 - 20 fps more in AA and AF.

This is very promising indeed, with Crossfire, you can still see a performance increase even at default settings at medium resolutions. This is noticeably different to SLI, at least based on our experience with a pair of GeForce 6600. With SLI, we mostly only see a significant increase in fps when AA and AF is enabled (when we're not CPU / system limited). However, that might not be entirely SLI's fault - it seems the Radeon X1600 vertex shaders are not as 'limited' as the GeForce 6600's. So, when you add more pixel processing power to your setup, in this case by putting in another card, the vertex shaders of the first Radeon X1600XT can still keep up with the demand of the pixel shaders units on both cards. We do not think this is a coincidence - ATI clearly designed the X1600 series to scale efficiently in a Crossfire setup. 

It will be very interesting to redo the SLI benchmark with new drivers or a new cards, possibly a pair of GeForce 7600 GS or GT. But for now, let's look at the rest of the benchmarks.


If you're playing this game with one (or two in Crossfire mode) Radeon X1600XTs, you should stick to 1024 x 768 to maintain a smooth frame rate, particularly if you're playing with AA and AF. At 1280 x 960, you're already dipping below the 30 fps mark at certain points and it's just gets worse at 1600 x 1200, with frame rates around 15 to 19 fps. That kind of frame rates can really be annoying, making the game feel sluggish and unresponsive. The minimum frame rates occurs mostly when you're looking at ships or objects at close up or have quite a large number of ships on screen.

By the looks of it, Crossfire isn't applying AFR in this game. So, either the embedded profile explicitly force default rendering mode or there's no profile for this game in the drivers. We think the former is true, since we can see that results with Catalyst AI enabled is slightly faster (though you won't notice it during gameplay), both at 1024 x 768 and 1280 x 960. We'll be looking at how to force AFR in this game, but for now let's accept the results as the default.

Crossfire isn't doing us any good here. While there are differences between the results (about 2 - 3 fps faster or slower) they're most likely caused by normal variations between runs. This may be disappointing, but it's certainly not surprising - we saw the same situation with SLI. Forcing SLI in multi card mode with this game actually made us experience lower frame rates.


While there's nothing to gain with a Crossfire setup (in default rendering mode). thankfully, there's also nothing lost in the frame rates department. The differences we're seeing in the results are very small, more likely normal variations between runs. Like Call of Duty, if you want to maintain a high minimum fps, better stick to 1024 x768. Frame rates near 30 fps can be very annoying in racing games, especially a fast paced one like this game.

This is probably one game that acts differently on Crossfire and SLI setups. We saw quite a significant increase in frame rates with AA and AF on an SLI setup (though at default settings we're actually getting lower frame rates). So, what's happening here? We saw no difference whether we're running with one card or two cards in this game or even with Catalyst AI enabled or disabled for that matter. Our guess is that there are no profiles embedded in the drivers for this game, so Crossfire is running in default mode. We're pretty sure it's not system related, after all we're seeing lower and lower frame rates as we move to higher resolutions, more so with AA and AF. Looks like this game is a good candidate to try to force AFR, after all we saw some improvements with SLI so Crossfire should be able to provide us with higher frame rates as well.


Hmm. Unlike the other games we tested so far, results with Catalyst AI enabled seems to be slower than with Catalyst AI disabled. The differences may be caused by normal variations between runs, but we doubt it. There's a clear trend here, notice the same level of frame rates from both single and multi configurations with Catalyst AI disabled. Obviously there's a game profile embedded on the drivers, however it's not optimal. We'll have a look later on trying to apply another game profile (and force AFR with Crossfire) with this game.

It's clear that the Radeon X1600XT (or two of them) is not fast enough to play this game at anything other than 1024 x 768, without AA and AF. Thankfully, you don't see these kinds of frame rates that often throughout the game. Now, if you look closely, you can see that in some cases, running with Crossfire can be slower - similar to what we saw with SLI. Hmmm, could this be true? It looks like Crossfire is not that much different from SLI - you can't really expect a performance increase in every game. We already saw that out of four games we tested so far, only one (Call of Duty) showed any frame rate increase with Crossfire. Maybe we spoke to soon when we say Crossfire has potential. How about newer games?


Crossfire can really help you get that extra fps in F.E.A.R, just don't expect too much from one or even two Radeon X1600XTs. There's no doubt, anything other than default settings at 1024 x 768 is way to slow for a single Radeon X1600XT. It's nice to know that having a pair of them in Crossfire AFR mode can help you reach playable frame rates, though it's still not fast enough to play with AA and AF enabled. Remember though, our test settings for this game are very high, not that much different from the maximum setting (no soft shadows, no volumetric lights). So all things considered, the results we're seeing here is not really that bad, not bad at all. Crossfire can mean the difference between unplayable and playable frame rates here, even if you're using a pair of low end cards. With high end cards, assuming the performance increase is present and similar, you probably will get playable frame rates at the highest setting with AA and AF enabled.


Wow! Looks like our Crossfire setup is really coming to its own now. You're not mistaken, those Crossfire results are nearly double the frame rates of a single card. Running at default settings, we nearly reach the limit of our Athlon 3500+ processor at 1024 x 768 - somewhere around 120 fps. While this is good news, do keep in mind that the difference we're seeing here is most likely not only caused by Crossfire alone, but also optimizations in Catalyst AI. Look at the single card results - you can see running with Catalyst AI enabled provide quite a bit of a boost in frame rates, though nothing as phenomenal as the results with Crossfire and Catalyst AI enabled. There's definitely a game profile for Quake 4 in the drivers.

Even accounting for Catalyst AI optimizations, the frame rates we gain with Crossfire is impressive - 20 to 30 fps both in default settings and with AA and AF. That's around 50 (default) to 80 (AA / AF) percent overall. This is great news indeed, Crossfire can provide you a significant boost in frame rates in newer games and this is with a low end card. With a high end card you'll likely be system limited, even with AA and AF. This is very different to SLI. If you look back at our SLI article, you'll see that Quake 4 is actually a bit slower with both a pair of GeForce 6600 or GeForce 7800GTX in SLI mode. Of course, that's with old drivers, but we don't have a pair of GeForce 6600 or 7600 right now, so we can't be 100 percent sure.

Now, let's look at the last two games - Serious Sam II and Splinter Cell: Chaos Theory. In addition to default and AA / AF results, we'll be looking at HDR results as well in these games.