Originally Posted by shadow001
Actually not quite all in fermi is improved over the HD5870,as the best case scenario for Fermi if it managed to hit 750 Mhz clock speeds,the floating point performance is 750 gigaflop second in dual precision math and tops out at 1.5 terraflops for single precision math.
HD 5870 cards are slower at dual precision(about 560 gigaflops),but way faster at single precision math,clocking in at 2.7 terraflops,and it's single precision math that used primarily in gaming scenarios,while dual precision is used for Cuda applications,specifically scientific work that requires the absolute maximum precision in the final result.....Keep in mind though that these are the theoretical maximums for each GPU.
Then there's the texturing abilities,with the HD 5870 using 80 texture units,while fermi seems to be using only 64 texture units,but as we don't know the individual capabilities on each unit in fermi,it's hard to make comparisons,but going on the rumored amount available in fermi and the clocks speed difference(200 Mhz faster for Cypress),ATI's HD5870 has more texturing power than Femi.
Lastly,even Anantech was pissed off in a recent article that clock speeds are still completely unknown,so it's hard to predict performance,and keeping clocks in the dark does very little to intimidate ATI,as they'll ship the refresh parts at the fastest clocks they can while still having good yeilds anyhow,as the point is to make money,not winning over Fermi with refresh parts so highly clocked that they'll have crappy yeilds as a consequence,and may the best win in the end.
Oh and as Xion mentioned,ATI's high end is the HD5970 cards,so double the above stats(1 terraflop dual precision,5 terraflops single precision,160 texture units),and Fermi is also down on memory bandwith too,so they have to be pretty carefull on the pricing of fermi cards here,as if it's close to the HD 5970,no one but the most hardcore Nvidia fan would buy one.