Multi-GPU graphics

In the past the only way to boost your PC's games performance significantly was to replace your CPU and graphics card. However, in 2004 Nvidia launched its SLI dual-graphics system, resurrecting an acronym and an idea that had been around since 1998. This time, however, SLI became central to Nvidia's chipsets and GPUs. ATI soon announced its own multi-graphics system, CrossFire, and over the past four years, both firms have put great effort into persuading consumers that the path to gaming nirvana is to use multiple graphics cards.

Initially, SLI and CrossFire were marketed as ways of linking two high-end graphics cards to play games at high resolutions and maximum detail settings. However, you can now double up nearly every type of graphics card, from £80 entry-level cards to £350 monsters with multiple GPUs and power demands that make shareholders of energy suppliers smile.

There are two markets for multi-GPU graphics: people who want a cost-effective way of boosting their PCs, and those who want the best possible gaming experience. If your PC is struggling to play the latest games, the question is no longer simply which graphics card you should buy, but whether you should buy a second graphics card or a single new one?'

While most articles have tried to answer this question by comparing the various merits of SLI and CrossFire systems, we believe that it's crucial to compare each variant of SLI and CrossFire with a single graphics card. After all, the question of whether SLI is better than CrossFire is academic if neither is significantly better than a single graphics card.

Over the next few pages we find out whether it's worth adding additional graphics cards to your PC, or if it's better to buy one powerful new card.

As the focus of this article is to find out whether adding an extra graphics card to your PC is a better upgrade than buying a single replacement card, we had to ensure that the results were as representative of a real gaming experience as possible.

Rather than relying on a synthetic benchmark such as 3DMark06, we measured the frame rate of each set of cards in real games. We did this by playing through a section of each of our test games and measuring the minimum and average frame rates using FRAPS (www.fraps.com). This way, the frame rates we recorded are representative of actual gameplay.

Most graphics cards reviewed in Custom PC are tested using three games (Call of Duty 4, Crysis and Need for Speed: Pro Street) at several resolutions, which allows you to choose the card that best matches your monitor. However, as SLI and CrossFire performance varies considerably between games, we decided to increase the number of games tested for this article. To do this we had to reduce the number of test resolutions to 1,680 x 1,050 and 1,920 x 1,200, but we feel that this small sacrifice is balanced by the larger number of games (ten as opposed to three), and the bigger picture they paint of the entire gaming market.

The games

We chose ten games: ArmA: Armed Assault; Call of Duty 4, Company of Heroes: Opposing Fronts, Crysis, Colin McRae: DiRT, Need for Speed: Pro Street, The Elder Scrolls IV: Oblivion, S.T.A.L.K.E.R., Supreme Commander and Team Fortress 2. They're popular games, and provide a good mix of first-person shooters, real-time strategy and driving genres. We planned to test Crysis and Company of Heroes: Opposing Fronts in both DirectX 9 and DirectX 10 modes, but even after several driver updates and patches, both games were unplayable in DirectX 10, with a minimum frame rate of less than 3fps.

We looked at several recent releases, such as Assassin's Creed and Frontlines: Fuel of War. However, like most console ports, Assassin's Creed runs very quickly on many recent graphics cards, while Frontlines: Fuel of War is unstable. We also considered using the incredibly popular World of Warcraft, but as it runs perfectly well on a three-year-old PC, it wouldn't reveal any differences in the hardware. We devised several benchmarks in Flight Sim X, but depending on the settings, the game either runs very slowly (less than 10fps) or fast (50fps or more) no matter which graphics cards are used.

The test rigs

One of the biggest problems when comparing multi-GPU systems is that SLI and CrossFire are incompatible with each other. For SLI you need a motherboard based on an Nvidia nForce chipset, while only certain Intel and AMD chipsets support CrossFire. However, with the recent release of the Intel D5400XS motherboard as part of the Skulltrail system, it's finally possible to run SLI and CrossFire on the same system. The board has two Nvidia nForce 100 chips for SLI compatability, and an Intel chipset for CrossFire. In other words, this is first article that can truly compare the differences between a single graphics card, CrossFire and SLI. We're certainly not recommending that every reader should build a Skulltrail system but, for the purposes of this review, it proved ideal.

This enabled us to test each graphics card on the D5400XS, and compare its frame rates with those from two-card SLI setups and two-, three- and four-card CrossFire systems. Sadly, 3-way SLI doesn't work on the D5400XS, so we tested this on an Asus Striker II Extreme motherboard based on Nvidia's nForce 790i Ultra SLI chipset. As no games support eight CPU cores, we ran the D5400XS with a single quad-core Intel Core 2 Extreme QX9775 CPU, as it's architecturally identical to the Core 2 Extreme QX9770 we installed in the Striker II Extreme. Both CPUs were overclocked from 3.2GHz to 3.6GHz, and each board had 4GB of RAM.

The reason for installing so much RAM is that Vista maps graphics card RAM to system memory, so it's important that you have sufficient memory free for games to run. For this reason we also conducted our tests on the 64-bit version of Vista. To ensure that even the most power-hungry cards received a stable supply of voltage and current, we fitted the test rig with the DXX version of the awesome 1kW Enermax Galaxy PSU. Although Nvidia and ATI release BETA drivers fairly often, we conducted each test using the latest WHQL-approved drivers to ensure system stability. All the ATI cards were tested using Catalyst 8.4, the GeForce 8-series cards using ForceWare 169.25 and the GeForce 9-series cards using ForceWare 174.74.

The results

You can find the test results on the graphs pages from p84, and a detailed explanation of how we tested each game. By looking at these figures, you can clearly see the resolutions and image quality settings at which each graphics card can run the games smoothly. The definition of 'smoothly' in this context is also important, as we consider both the minimum and the average frame rate from our tests. This is because the eye can perceive the lowest frame rate, so a low minimum will mean that the game stutters and jerks even if the average frame is high. For this reason, we only define a game as playable if the minimum frame rate is 25fps or higher.

In addition to the 1,040 benchmark results, on each review page you'll find a percentage score for performance in the three game genres (FPS, RTS and Driving), plus an 'all games' percentage calculated as an average of the minimum frame rate in all ten games. For the two-card multi-GPU systems, these numbers indicate how much faster or slower this configuration is than a single graphics card. As you're unlikely to be considering a three- or four-card multi-GPU system unless you already have, or have considered buying a two-card setup, the percentages for these reviews are relative to a two- or three-card setup respectively. This means that whatever multi-GPU configuration you're interested in, you can immediately see the performance that adding extra cards delivers.

As well as measuring the performance of each card in our test games, we measured the peak power consumed by our test PC while running 3DMark06. This enables us to compare the power efficiency of the cards. This graph is shown on p98.

Custom PC would like to thank Crucial, Intel, Sapphire and Scan for their help with this article.