A little while ago Rage3D, along with eight other websites from around the World, kicked off a new collaborative series of articles designed to examine the multi-GPU technologies offered from graphics heavyweights ATI and NVIDIA.

The series is broad in scope and innovative in design. Six separate systems, three each based on Crossfire and SLI technologies, were assembled to represent three budget levels a prospective user might be faced with: a low-budget level $1,000 plateau, a gamers $2,000 plateau, and an Unlimited plateau where we throw all the rules out the window.

• Josh over at Penstar Systems kicked things off with an Introduction to the series and detailed which sites were participating and how the whole thing was conceived and how it would work.  Josh took the reins as the series editor-in-chief, so he didn’t get any fancy hardware to play with and had to deal with all the logistical nightmares (and there were a lot of them). Sucks to be Josh!

• Following that, Tim Smalley at bit-tech.net took us on a tour of Crossfire and SLI, the multi-GPU technologies from ATI and NVIDIA respectively, and the graphics hardware that would power the six separate systems we would examine in the series. Tim had to be the first out of the gate with his piece so the pressure was on. Luckily for us he also had the resources and contacts needed to acquire much of the hardware distributed to the various sites. Cheers to Tim.

• A few days after Tim’s piece, Bjorn of Bjorn3D kicked in with an extensive look at the Crossfire and SLI motherboards that would form the basis of the systems used in the series. Bjorn managed to do a bang-up job with his piece and provided the first benchmarks in the series.

• Following up on that was Jon at HARDiINFO with the highly anticipated “Common Benchmarks” portion of the series. Jon provided a deluge of graphics hardware results for us to pour over and started us in on the real meat of the series. Poor Jon also had to deal with the first bit of controversy in the series thanks to some pesky Control Panel setting errors. Reviewers everywhere will tell you that a complete lack of sleep will often bring on such controversial errors, and now Jon has firsthand experience.

• Subsequent to Jon’s article was Howard's over at Neoseeker. Howard, who has the distinct pleasure of being able to call yours truly a fellow Canadian (even though he’s living on the wrong end of the country), was tasked with presenting benchmark results from a list of best selling games as provided by Xfire (not to be confused with Crossfire!).

• Mike over at nV News chipped in next with his look at a series of games that aren’t normally found in many reviews… uncommon  games, if you will. Mike also took the opportunity to start testing Image Quality differences with a look at Anti-Aliasing image quality. Frankly we should have allocated an entire article toward image quality. Live and learn.

• Next up was Brian over at Legit Reviews who graciously swapped positions in the schedule with me after I ran into some hardware problems. Brian's job, like Mike's and mine, was to test more of those Uncommon games.

And thus is Rage3D’s turn at the plate. My job today is to present you, fine reader, with benchmark results from another list of games that you don’t normally find in a normal graphics card review. I will start out by prefacing the whole shebang with a simple warning: my library is not unlimited, and not all games can be benchmarked within the bounds of time and reason. This warning will serve you well if you come to find yourself questioning my particular game choices for this part of the series. I did, however, try and keep to the whole “sim” theme, since that is my personal favorite genre.

I will also take over from where Mike left off with his Anti-Aliasing image quality comparison and do some texture filtering quality tests. Texture filtering makes the biggest impact on overall image quality if you ask me, and I think many would agree that ATI and NVIDIA go about it very differently and produce very different results, so it’s important that we take a close look at it.

Before we get started Rage3D would like to express our sincere thanks to our sponsors, without whose help this series of articles would not have been possible:. abit for their generosity in providing many of motherboards used in the series of articles, AMD for the all the Athlon CPUs, ATI for the videocards for our European friends, BFG Tech for sending out bucket loads of their videocards, ECS for early access to their Socket AM2 ATI motherboard, NVIDIA for coming up with the idea and helping get much of the hardware, Sapphire for their videocards, and SuperTalent for providing their kickass DDR2 memory kits for use on the AM2 platforms.

Let’s get to it.

---

« Previous in the Multi-GPU World Tour: Uncommon Benchmarks at Legit Reviews

Rage3D doesn’t necessarily subscribe to the “real-world” method of benchmarking graphics hardware. Instead, we think it’s important to reduce the impact other system components have on the graphics hardware as much as possible. This gives the graphics hardware some breathing room and allows the reader to gain some perspective on what it is capable of.

However, since in this article we are testing entire platforms and not necessarily individual components like graphics cards, we’ve decided to take a different approach and have made some changes to our normal benchmarking methodology.

First of all, all of the test results you will see in this article were produced using Fraps to record frame-rates for each setting and resolution. In most cases, resolution and setting scores are an average of several passes at that resolution and setting. The number of passes depended on the reported frame-rate variation for each game; the higher the variation, the more passes we would make.

Some games, such as Pacific Fighters, have accurate track recording and replaying capability and are very reliable when benchmarked with Fraps. In those cases, we generally made only two passes to ensure that no anomalies existed. Other games, such as Colin McRae Rally 2005, which have no replay capability and require us to manually play the game, were tested five times per resolution and setting. The lowest and highest scores were then discarded and an average of the remaining three gave us a final score.

Each pass was as close to what you would normally experience in a normal gaming session while at the same time keeping the game-play as repeatable as possible. This usually meant taking things a little slower around the corners in the racing sims to avoid any unrepeatable mistakes, not engaging difficult enemies and flying the same routes in the flight sims, and taking the same actions in the military squad based sim.

To get realistic results for each platform, an average best fame-rate per game was determined and in-game settings were adjusted to reach that frame-rate. This doesn’t mean that we adjusted settings per resolution but rather a resolution was picked that we thought most gamers would be using if their display matched the platform*, playable frame-rates were obtained at that resolution (“playable” is a purely subjective where frame-rate is concerned, but that’s entirely unavoidable), then we used those same settings throughout the resolution range. The settings used were determined on the ATI test system and kept consistent on the NVIDIA system it was being compared with (ATI was used as the base system simply because the $2,000 ATI system was the first one we could put together at the time. We then decided to stay consistent with the other system levels and use ATI as the base there as well). This keeps the ideals of the platform tests intact, while still allowing us to see how well the graphics hardware scales over resolutions.

While this approach won’t give you the exact performance characteristics of any one game on your system (there are simply far too many variables for us to be able to realistically claim otherwise), it should be able to accurately tell us which platform is performing best.

For the NVIDIA systems, their graphics Driver control panel graphics settings were adjusted on a platform by platform basis. Simply put, this means that for the $1,000 platform, the drivers were left at their default values, for the $2,000 platform the system was tested at both the default and highest image quality, and for the Unlimited platform high image quality was used exclusively. For ATI hardware, the default control panel settings were used throughout each platform.

This might not seem fair at first, to adjust the image quality on the NVIDIA platform while leaving the ATI platforms at their default values, but you will understand why we think this is important when you check out our image quality tests on the next page.

*generally this meant 1280x1024 on the $1,000 platform, a balance of 1280x1024 and 1600x1200 on the $2,000 platform, and 1920x1200 or higher on the unlimited platform.

 

System Specifications

$1,000 Platform	ATI	NVIDIA
Motherboard ( chipset )	ABIT AT8 ( ATI RD480/ULI m1575 )	ASUS A8N-SLI Premium ( NForce 4 SLI )
CPU ( Socket )	AMD Athlon X2 3800+ ( Socket 939 )
Graphics Card ( Drivers )	Sapphire X1600 Pro x2 ( Catalyst 6.6 WHQL )	BFG 7600 GS OC x2 ( Forceware 91.33 Beta )
Memory ( Timings )	1GB (2x512MB) Samsung PC-3200 DDR ( 2.5-3-3-7 )

$2,000 Platform	ATI	NVIDIA
Motherboard ( chipset )	ABIT AT8-32X ( ATI RD580/ULI m1575 )	ABIT AN8-32X ( NForce 4 SLI x16 )
CPU ( Socket )	AMD Athlon X2 4400+ ( Socket 939 )
Graphics Card ( Drivers )	Sapphire X1900 GT x2 ( nforce4_hotfix )	NVIDIA 7950 GX2 ( Forceware 91.33 Beta )
Memory ( Timings )	2GB (2x1024MB) Corsair PC-3500 DDR ( 2-3-2-6 )

Unlimited Platform	ATI	NVIDIA
Motherboard ( chipset )	MSI K9A Platinum (ATI RD580/SB600)	Foxconn C51XEM2AA (NForce 590 SLI)
CPU ( Socket )	AMD Athlon FX-62 ( Socket AM2 )
Graphics Card ( Drivers )	ATI X1900 Crossfire/X1900 XTX ( Catalyst 6.6 WHQL )	BFG 7900 GTX OC x2 ( Forceware 91.33 Beta )
Memory ( Timings )	2GB (2x1024MB) SuperTalent PC2-6400 DDR2 ( 4-4-3-8 )

Components common to all systems were a Silverstone Strider ST56F 560W PSU (Crossfire and SLI Certified), 160GB Seagate Barracuda 7900.9 SATA HDD, and Windows XP SP2 with the latest updates and patches.

If you’ve not read the previous articles in the series I suggest you go back and check them out. They contain a thorough rundown of most of the graphics hardware and motherboard components use in this article.

Game Benchmarks

• GTR2, demo version
• Ghost Recon: Advanced Warfighter, v1.20
• GT Legends, v1.1.0.0
• Pacific Fighters, v4.04
• Richard Burns Rally, v1.02
• Lock On: Modern Air Combat, v1.02
• Colin McRae Rally 2005, v1.0

Amongst those who follow this industry, there’s nothing much more controversial than texture filtering. It’s that one area where both ATI and NVIDIA take a lot of liberties with image quality in order to improve their overall performance.

Both companies now include angle-dependant optimizations that reduce filtering quality where they think it’s not all that important. ATI introduced angle-dependant texture filtering to the world when they released the Radeon 9700 Pro in the summer of 2002. At the time NVIDIA’s Geforce 4 series was their flagship part and offered true, angle independent texture filtering that looked great, but could not compete with ATI’s optimizations in terms of performance.

Unhappy with losing to ATI in many performance tests, NVIDIA introduced their Geforce 5 series which featured angle-dependant filtering technology of their own, and took things a few steps further than ATI had gone.

Notoriously known as “brilinear” filtering, NVIDIA’s driver default mip-map blending quality fell somewhere between bilinear and full on trilinear filtering. It doesn’t look bad in in-game screenshots, but during game-play you can definitely notice the textures shimmering where they transition from higher to lower resolution.

The easiest way to get rid of the shimmering effect is to change the driver control panel image quality setting from Quality to High Quality. Doing that turns off some of NVIDIA’s filtering optimizations and reduces the shimmering effect substantially. Changing this setting brings NVIDIA’s texturing filtering quality pretty much on the same level as ATI’s, as you can see in the AF Tester screenshots below.

However, ATI, taking advantage of immense negative press surrounding NVIDIA’s filtering algorithms last summer, raised the bar even further with the introduction of “High Quality Anisotropic Filtering” when they launched the R520 based Radeon X1800XT in the fall of 2005. This feature removes the angle-dependant filtering techniques that were introduced in 2002 and brings us back once again full circle to high quality texture filtering.

While screenshots are far from the best way to demonstrate texture filtering quality, the following sets will hopefully serve to illustrate the rather blunt point made above. Click the thumbnails to launch a new interactive page containg high-res, PNG-24 images.

In game these difference are very noticeable, and in some cases the differences are noticeable even in still screenshots.

If you go back to Mike’s article and look at the edge anti-aliasing you will see that not only does ATI offer better texture filtering, but they also offer better edge AA as well. With these advantages, there is little doubt that ATI offers better overall image quality.

Due to some unavoidable hardware problems with the Foxconn C51XEM2AA motherboard we've had to delay this section of the article. We will have it up as soon as possible, please stay tuned!

Though not interesting without the 7900 GTX to compare with, here regardless are the results from the X1900 Crossfire tests.

NVIDIA’s SLI is clearly more stable and mature than ATI’s Crossfire, there is no debating that fact. I never experienced any game crashes or rendering errors with SLI, and the one game I was expecting to have some problems with (Pacific Fighters) was apparently recently fixed and now works and scales fine. With Crossfire not only did most games not scale, but I actually had to disable Crossfire just to get them to launch. Elegant degradation is not in ATI’s vocabulary, apparently.

Not only is stability on NVIDIA’s side, so is ease of use and practicality. There is no master/slave stuff and there is no giant pain in the ass dongle with thumb screws that break off and are way too small and cables that are way too stiff to bend nicely like the monstrosity you get with hardware Crossfire. NVIDIA’s SLI bridge-board is elegant, easy to use, and doesn’t get in your way. Granted you can’t defend yourself against hordes of evil undead zombies with it like you can ATI’s dongle, but really, who wants zombie goo all over their computer anyway.

I remember last year as ATI was announcing Crossfire all the hype they were making about not needing to use game profiles and that every game “just worked”. A shot at NVIDIA and their SLI profile system, clearly. Turns out ATI uses profiles in much the same way as NVIDIA. The only difference is that ATI has far fewer of them (so far fewer games actually work with Crossfire compared to SLI), and they are closed up in an encrypted .DLL or some silliness so users can’t modify it and add their own profiles (NVIDIA’s are in a plain text .XML file, easily modifiable).  With ATI’s software designers all neatly tied up with Vista right now, I wouldn’t hold my breath waiting for user customizable profiles for Win XP.

Speaking of profiles I noticed that the new SimBin racer GTR2, even though it’s not slated for retail release until late September, is already in NVIDIA’s profiles. Meanwhile, even though ATI releases drivers once a month, it can sometimes take quite a while for them to get a profile in their drivers to get it working with Crossfire. And that’s for high-profile games too, I won’t be surprised if lower profile “niche” games like GTR2 never get added.

Now that I’ve ripped ATI a nice new one, I’d like to start in on NVIDIA. There’s nothing bad I can say about SLI, it works and works great, but NVIDIA really needs to do something about their image quality. Their Anti-Aliasing quality lags behind that of ATI’s in just about every case, but of particular inferiority is their texture filtering.

I can reason out the AA complexities, but why is it NVIDIA feels they need to resort to absurd filtering schemes like “brilinear” filtering to gain a few extra frames per second is beyond me. You get a shiny new Geforce and expect all the nice SM3.0 eye-candy, only to have it all ruined with this brilinear crap.

Not only does it make them look bad (literally and figuratively), it also makes them look like they are desperate for those extra frames when clearly they shouldn’t be; their cards are more than powerful enough to not only offer trilinear filtering, but also powerful enough to let them go back to true angle-independent filtering as well. They build themselves up a significant technological lead with SLI, and then they go and ruin it with intentionally inferior image quality. It’s a downright shame.

So, what was apparently so easily cut and dry a few pages ago now has a little caveat attached to it: speed or IQ. I suppose, though, that since you’re reading this you’re probably interested in a Multi-GPU system, and even with the image quality advantages ATI has, it is nothing near enough to make up for the vast stability and usability disadvantages Crossfire has compared to SLI.

---

Next in the Multi-GPU World Tour: Conclusion at Guru3D »