There's only been one hardware review site (after checking more than 20) that directly compares a 5870 against 2x 4890's in CF:
Many thanks to Firingsquad, an excellent review site (
http://www.firingsquad.com/hardware/ati_radeon_hd_5870_performance_preview), here are the benchmark snapshots from FS.. Corei7 X975 @ 3.33 GHz was used, but Crossfire has additional CPU overhead, so CF should be limited by the CPU before anything else, and 2560x1600 will be chosen to ensure that the CPU bottleneck is avoided in as many games as possible. 2560x1600 is also a good indicator of upcoming games next year, as some of the games out right now are not so demanding on those cards.
A 5870 does 72.3 fps
A 4870X2 does 82.5 fps
2x 4890's CF does 85.8 fps (18.7% increase over a 5870)
A 5870 does 58.6 fps
A 4870X2 does 61.7 fps
2x 4890's CF does 72.4 fps (23.5% increase over a 5870)
A 4870X2 does 31.3 fps
A 5870 does 33.8 fps
2x 4890's CF does 37.1 fps (18.5% increase over a 5870)
A 4870X2 does 20.2 fps
A 5870 does 20.6 fps
2x 4890's CF does 23.6 fps (16.8% increase over a 5870)
A 5870 does 47.8 fps
A 4870X2 does 50.6 fps
2x 4890's CF does 61.2 fps (28% increase over a 5870)
A 5870 does 32.1 fps
A 4870X2 does 35 fps
2x 4890's CF does 39.4 fps (18.7% increase over a 5870)
A 5870 does 63 fps
A 4870X2 does 63.2 fps
2x 4890's CF does 73.8 fps (22.7% increase over a 5870)
A 5870 does 72.6 fps
A 4870X2 does 88.9 fps
2x 4890's CF does 101 fps (39.1% increase over a 5870)
A 5870 does 72.1 fps
A 4870X2 does 79.2 fps
2x 4890's CF does 92.7 fps (28.6% increase over a 5870)
A 5870 does 43.8 fps
A 4870X2 does 49.5 fps
2x 4890's CF does 58.3 fps (33.1% increase over a 5870)
A 5870 does 55 fps
A 4870X2 does 74 fps
2x 4890's CF does 89 fps (61.8% increase over a 5870)!!!!!
A 4870X2 does 52.3 fps
A 5870 does 54.6 fps
2x 4890's CF does 61.3 fps (17.2% increase over a 5870)
The results are also quite similar with 8x AA instead of 4x AA, where 2x 4890's remain the undisputed king over a 5870. A 5870 does gain a 2% advantage overall when using 8x AA, against a 4870X2, but it is still not enough to beat a 4870X2 overall.
In every single game of the benchmark test suite done by Firingsquad, a 5870 has lost to 2x 4890's in CF. Sometimes, 2x 4890's in CF beats out a 5870 by more than 50-60%.
Are the drivers to be blamed? Many would like to hope for a miracle boost in performance from driver optimizations, which has never been done by ATI before to the point where there's a 15% increase across the board. It has been done by Nvidia a couple times in the past several years, in which a new driver set brought about 10-20% increase in performance in a handful of games. Usually, it was after optimizing for a new GPU architecture or for new graphical features in games. However, in the case of 4890's in Crossfire and a 5870, both are very similar architectures, if not the same minus DX11.
A 5870 chip actually looks like two 4890's infused into one chip, with two halves of 800 shader units on both sides. In theory, it should perform identically to two 4890's, according to the identical 850 MHz clock speed and exactly 2x the quantities of shader and texture management units and render back-ends (ROP's). Actually, in theory, a 5870 should perform better as it removes the dependence upon Crossfire.
Let's see if 2x 4890's would no longer beat a 5870 in all of the above games if and only if the total memory bandwidth of both cards would be exactly the same as that of a 5870.
That would mean downclocking the memory all the way down from 3.9GHz to 2.4GHz effective for each 4890 card, so that the total bandwidth matches that of a 5870 card with 4.8GHz memory.
Only then can we know for sure...
EDIT:
A grand hypothesis: I expect that two 4890's in crossfire will not perform any better than a 5870 in any of the above games if the total memory bandwidth is reduced to that of a 5870. However, the performance would be within 96-100% of a 5870 in several games.