Radeon Catalyst 13.8 Beta Improves CrossfireX
In short, the latest Catalyst 13.8 Beta driver should bring a tremendous improvement to the quality of gameplay on CrossfireX (multi-GPU) configurations of HD 7000, HD 6000 and HD 5000 series GPUs. Benchmarks show the reduced differences between frametimes in CrossfireX, compared to previous drivers. Nvidia's SLI still provides even better results, but the gap is not enormous anymore. The new Frame Pacing so far applies only to DX10 and DX11 games on a single display, with the support for other display configurations and DirectX and OpenGL versions coming later. There is also some long history showing why this is important news for the AMD GPUs.
When it comes to multi-GPU setups, all the details about using several GPUs to render the same game open a big can of worms. There are a lot of issues that can arise both for performance and gameplay experience. Nvidia and ATI went in different directions in the way their multi-GPU solutions were set up. Nvidia enforced very tight conditions on their SLI solution, requiring the same GPUs to be used, as well as restricting it to motherboards with Nvidia chipsets (before Intel Nehalem arrived in 2008). ATI, on the other hand, attempted to make their approach more flexible, allowing to pair similar cards (but not identical) in Crossfire (like 4850 and 4870). After acquisition into AMD, Hybrid CrossFireX allowed a dedicated AMD GPU to work together with the integrated AMD GPU.
Before AMD's Radeon HD 6000 series, there was little question about the performance of the two competing solutions. Nvidia's SLI was superior in most cases, even allowing a pair of slower Nvidia GPUs to beat a pair of stronger AMD GPUs. HD6000 changed that, even giving a small edge to the CrossFireX in benchmark results. This was only small, but even in cases where benchmarks gave higher numbers to AMD GPUs, there were reports from people who experienced a smoother gameplay from Nvidia's cards, especially in case of multi-GPU setups. There was also common knowledge about micro-stuttering on multi-GPU setups, but based mostly on subjective experience, with little hard evidence. Most reviews analysed only the average framerates with FRAPS, sometimes looking at minimal framerates as well. Some sites, like HardOCP were already showing detailed framerate graphs with some subjective evaluation, but overall there was little hard data for a detailed look at how smoothly games are rendered. Average FPS comparisons are good only as long as there are not any other issues.
The very loud call to change the way graphics cards are benchmarked came in late 2011 from The Tech Report with their article Inside the second. The average FPS was out of the spotlight, the time a system took to render each frame was more important. The data was still collected with FRAPS alone, which missed some parts of graphics pipelines, but still provided a lot of information. Those results vindicated some earlier claims – there was a real stutter in multi-GPU setups, and it was worse on AMD CrossFireX. However, it was not the whole picture, due to the things between FRAPS measurement and display of the frames, which affects both the lag between game and controls and the smoothness of final output.
Due to the complexity of modern GPU hardware, a lot rests on the drivers. The way they work determines how much performance cards can provide and how smoothly the rendering goes. Since their release, both HD 7000 and GTX 600 graphics cards were getting constant improvements with new driver versions. The new benchmarking methods found some problems with frame delivery on Radeon GPUs, which were later fixed by AMD with newer drivers. Still, in 2012 the details about multi-GPU performance remained shady without a precise way to measure the frames shown on the display. Fortunately, the tools to get those measurements were already in development.
The idea was quite simple in principle – with no V-sync or triple-buffering enabled, a GPU can start drawing a new frame without finishing a previous one on top of each other. The image shown on screen will consist of several rendered frames. It is known as screen tearing, not nice during gameplay, but turned out to be very useful for benchmarking. By looking at all the images on monitors, the amount of frames rendered, and how long each of those could be seen can be calculated. Even on a 60 Hz monitor or capture card, framerates in hundreds of FPS can be measured as long as you know which frame is which. Several review sites were developing such methods, but the version which became standard came this year from Nvidia – FCAT.
There is definitely a reason to be careful, as it would be convenient for Nvidia to bash its competitor silently. Fortunately the basic software building blocks are open to checks and outside tampering and there does not seem to be any malicious code inside. On software side there is an overlay for 3D games, which adds coloured stripes to each frame to help identify it. After the capture, the analyser tool extracts the information about frame times. In between, a beefy hardware is required for a correct capture. To benchmark at 2560x1440@60Hz a very capable capture card is required as well as a RAID array of fast SSDs to handle the 500 MB/s of video data. These requirements, as well as the additional time needed to analyse data, mean that only some hardware review sites can manage to do such tests. Some of the ones that did, have found out that Nvidia adds some lag to SLI, but in turn gets very smooth SLI experiences. In CrossFireX case, FRAPS was giving compliments, as the display side was much worse.
AMD has also taken notice of those FCAT results and have promised to improve their multi-GPU drivers. While Radeon HD 7990 was released with only minimal improvement in smoothness, the latest 13.8 Beta definitely shows that AMD is planning to keep its word. It is also nice, that these improvements should also help older HD 6000 and HD 5000 series cards. Considering the use of HD 7970 in CrossfireX among those playing at high resolutions, the driver improvements are better late than never.