If the nVidia Tesla M2090 can only do 1.3TeraFlops in single precision, and the AMD HD6990 can do 5.4TeraFlops, how come the Tesla costs substantially more? The HD6990 also does better in double precision. Anyone care to answer? I know, it's pretty technical.
The Tesla isnt even a workstation class card, its a super computer class card, meant for GPU compute tasks with multiple tesla cards running together. It uses ECC memory and has much more data checking and better tested drivers than a desktop level card does. It also costs a lot more because nVidia will provide optimized drivers for certain programs and assist in creating custom drivers for the purpose.
Its also important to note that once you get into direct compute tasks that sometimes an nVidia card is better and sometimes an AMD card is better, if its a massive quantity of simple things repeated over and over the extra SPs of an AMD card will do better, but if its a limited number of more complicated tasks the better serial performance of a cuda core will be more useful as the extra cores of the radeon will be sitting empty often.