SLI vs. CrossFire, Part 2 – High-end multi-GPU scaling

Conclusion

This has been quite an enjoyable hands on experience with these top video cards evaluating SLI and CrossFire video card combinations. Although there is much left to cover, we have generally observed good multi-GPU scaling that was generally more noticeable at higher resolutions; and of course, we want to see what over- and under-clocking our CPU and GPUs may do.

Of course, this is part 2 of our ongoing CrossFire versus SLI series. There are going to be at least two or three more parts that will delve into CPU and GPU overclocking and the effects of difference CPU platforms and architectures on performance. Eventually we want to set up testing of “micro stutter” as observed in multi-GPU setups which is easy to see for yourself but very difficult to demonstrate in an evaluation.  We have a digital camera capable of 240 fields per second and will attempt to make videos for you.

Micro Stutter and Greater than 100% Scaling

Micro stutter bothers some gamers so much so that they cannot ignore it and they return to using a single powerful card while others never notice it at all and will swear that it does not exist. Micro stutter does exist as it is inherent to Alternate Frame Rendering (AFR), but drivers minimize its effects so that generally the performance gained by the second card outweighs any disadvantages.

We also note that in many situations where micro stuttering is evident, the CPU or Graphics are straining and often it may be minimized somewhat by lowering settings.  If you would like to weigh on our future testing methods or to specifically discuss micro stutter testing methods, feel free to comment below, ask questions, make requests, or have a detailed discussion in our ABT forum.

We have also occasionally noted CrossFire and SLI scaling greater than 100%. Often this can be attributed to “benchmark noise”, the way averages are rounded off, or the nature of Alternate Frame Rendering (AFR) where multi-GPU scaling may be slightly better than 100% due to the frames that are read ahead by the CPU.

If frames for a multi-GPU system are only pre-rendered as for a single card there would be less than optimal scaling. The AFR dual-GPU system likely has to render ahead further than a single-GPU system does because there are now two GPUs to feed. In this scenario, the framerate may possibly have higher than 100% scaling during some GPU-bound situations.

In other words, a GPU can render only one frame at a time. If there were no pre-rendering at all, the CPU has to wait before constructing a new frame, while pre-rendering allows the CPU to keep on preparing frames while the GPU is busy. Since the prepared frames are stored in an offline buffer, the GPU ideally never has to wait for the CPU to construct them. Now since the pre-render value is likely higher for a multi-GPU system than it is for a single GPU system, it is possible to achieve greater than 100% scaling. The multi-GPU system’s higher pre-render values makes it probable that the CPU can continue working on frames instead of stopping for a single video card with a lower pre-render value. Of course, this would cause more input lag but that is beyond the scope of this discussion.

Other factors can also contribute to greater than 100% scaling which include “benchmarking noise” – the margin of error built into benching is generally accepted as 1-3% – and also driver bugs that might hold back a single GPU but which don’t show up in multi-GPU situations. Usually you will not encounter this; ideally scaling approaches 100% if your SLI or CrossFire drivers are working perfectly.

Continued Investigation

Hopefully, the benefits and disadvantages of CrossFire or SLI will become even clearer as we continue on in our investigation of SLI versus CrossFire. The top cards from both AMD and Nvidia scale very well in newer games and in most of our 22 games that we use as our regular benchmark suite.

One thing to note is that all of our video cards were run at their stock speeds.  For our next Part 3 evaluation, we will attempt to overclock our HD 6990 so that it approaches or matches the stock speed of the HD 6970.  We will also overclock our GTX 590 to see how it scales as its two cores are basically underclocked GTX 580s much the same as the HD 6990 has two 6970 cores on a single PCB.  Changing the clock speeds will give us a very good idea of further scaling of CrossFire-X3.  And of course, we will over- and under-clock our CPUs to see if there is a “bottleneck”.

We will not give any final judgments now but will withhold them until after we are done with this series. We hope to have a new powerful flagship FX series CPU coming from AMD and we hope to give you a hands on full evaluation as soon as it is available. Of course, it will fit in very well with this series. We are in the middle of one of the most interesting times in the recent history of gaming as AMD and Nvidia are locked into a battle to win your hearts and money.

Stay tuned, there is a lot coming from us at ABT including a Catalyst 11.9 Performance evaluation this weekend – and of course, the next part in our ongoing series – CrossFire vs. SLI – Part III.  Stay tuned!

Mark Poppin

ABT Senior Editor

Please join us in our Forums

Become a Fan on Facebook

Follow us on Twitter

For the latest updates from ABT, please join our RSS News Feed

Join our Distributed Computing teams

• Folding@Home – Team AlienBabelTech – 164304
• SETI@Home – Team AlienBabelTech – 138705
• World Community Grid – Team AlienBabelTech