Well, I have some questions that have been bothering me for some time. Thought I’d see what everyone thinks.
Primarily I am curious as to how crossfire works from a technical standpoint. The common wisdom is that the fastest card will be underclocked to run at the same speed as the slower card. Yet I have it on good authority that this is not the case, and that the cards are actually running asynchronously.
I have no real sources or proof on this other than the mentioning of a mod at the ati forums. Though I believe what he says.
I’ve always thought that it was asynchronous; at the least CCC/gpuz will never show any downclocking from crossfire.
In the above review the 4850 crossfire can be seen to perform almost exactly the same as the 4850 plus 4870. Indicating that if they are actually running asynchronously that they are not running it very well.
If anyone can shed some light on the workings of crossfire I am quite interested in it.
That being said, I tried to figure this out by over-clocking and under-clocking my 4890’s before. The results were very inconclusive. I found that the performance was more limited by the slowest of the pair. I found that I generally got better performance out of a small over-clock on both than a huge over-clock on one. Though the differences were small enough I didn’t consider them viable, as they were not enough above random variance for me to pay them any mind.
The possibilities I see it would be:
1) Crossfire does work asynchronously (Each card is given a different load based on its speed), but this is determined by pre chosen values in the drivers which means the load difference is so small as to not make much difference at all.
2) The drivers just don’t work very well in mixed crossfire set-ups, for reasons I’d like to figure out. (Perhaps the pre determined asynchronous values do not take into account over-clocks strictly limiting the benefit of a gap I speeds across the cards)
3) The cards do actually operate at the speed of the slowest but this is hidden from the end user through unknown means. (It is blindingly obvious, and one of the few things I can test without issue, that if one card is over-clocked and one is not GPUz will report the selected speed on each card regardless of load, crossfire, etc.)