China successfully developed "Milky Way One" Supercomputer

[blob]

i would say its dependent on the application. video encoding works on ATi hardware much better but good software isnt really there yet. ATi certainly has a unique advantage in double precision in that it can pair 2 single precision units into one DP unit. they dont waste any die space doing that. nvidia cant do that. they have to dedicate die space for DP.

the current f@h ATi client has been tainted by r600. the entire ATi implementation is a mess. the two main theings holding it back is the fact there is no cache in r600 and rv770 and up have cache but its not used. the speed would easily double if they utilized it. the other bottleneck is the cpu. it has to be used for random number generation. this is why the 5870 has the same ppd ad the 4870. but dont worry gpu3 will fix all of those problems. it uses opencl too. i would expect improvements in the nvidia client too. i would not blame all of the problems on stanford molecular dynamics is really complicated and nvidia is taking a smart approach and helping programmers with libraries and optimizations. ATi has not really done this well. the reason why companies in the past with massively parallel architectures failed is usually because of software.

Maybe you should checkout Milkyway@home and Collatz Conjecture

Quote by Andreas Przystawik, programmer who optimized the Milkyway@Home client for GPU Computing:
I have no idea what Folding is doing in its code. But you may want to look at two other distributed computing projects using ATI cards: Milkyway@home and Collatz Conjecture.

Both scale virtually perfectly with the higher shader count. That a HD5870 achieves this with the Collatz Conjecture project is also showing that ATI has done a tremendeous job with the memory controller efficiency as it is quite bandwidth hungry.

And I guess I don't have to tell you how pathetic a GTX285 looks compared to a HD5870 at the Milkyway project which uses double precision calculations (it barely matches the performance of an old and rusty HD3850). Nvidia badly needs a factor 8 improvement here to even approach a Cypress, let alone to pass it. And by the way, an nvidia engineer helped the project with its CUDA application, I guess one can assume it's near the maximum performance the hardware is capable of.
I have high hopes for Fermi, but frankly I still doubt a bit it will pass Cypress on pure number crunching.

Oh, by the way, both projects use the Brook+ layer as a base for the ATI applications. So that has nothing to do with the possible scaling.