Who Scales: How Often?

A major topic in the multiGPU arena is software support. And there are two large factors here: how many titles benefit and how much do those titles benefit. In the past we've seen SLI provide scaling more frequently and consistently than CrossFire (especially right when games come out). With CrossFire we'll often see support for older games get broken in newer drivers and then fixed when a review site happens to stumble upon the issue. But we've also noted that when CrossFire worked, it worked really well. It's honestly been a long time since we did a quantitative analysis of how SLI and CrossFire really stack up as technologies, and there's no time like the present.

First we will explore whether performance scaling happened in our suite of games. We've looked at two different metrics to judge our cards, both of which look at percent increase from 2 GPUs. If we consider the success of a multiGPU solution to be contingent on a performance improvement of at least 30% out of a possible 100%, we can count the number of times we see success in our benchmarks as a benchmark. We ran 21 different tests (7 games at 3 different resolutions), so keep that in mind when looking at this of list successes per configuration.

Since this takes into account CPU limited cases, our higher performance SLI and CrossFire solutions will see cases where 1680x1050 or even 1920x1200 isn't a high enough resolution to allow for any real improvement. Cards that look good by this metric are ones that both scale well and start off at a low enough performance point so as to allow good scaling to happen even at lower resolutions (well, lower for multiGPU application anyway). This shows the GTX 260 and the 4850 hit a sweet spot in terms of scaling and baseline performance in modern games to provide benefit for a larger number of users (many more people have 1680x1050 and 1920x1200 than 2560x1600 monitors). Because this 9800 GTX+ is older, we see headroom here too.

If we exclude the simply CPU limited cases and look at cases where the multiGPU solution got near zero or negative performance improvement we see a slightly different picture. Our data is on a per game basis, so all of these numbers are out of 7.

This shows cases where certain multiGPU configurations have zero value to help improve performance because of some failing of the graphics solution. All these cases happen to be issues at 2560x1600 where the resolution proved too much to handle because of the limited amount of onboard RAM.

It's also important to point out that the Sapphire 4850 X2 doesn't suffer from the problems of the 4850 CrossFire we show here. The Sapphire card scales and performs well in every test we ran.