Much ink and little info. Please, improve your writing skills. It shows that you have had no training, or no interest on learning on your own how to convey information in written media. You wasted my time, and time is the most valuable asset a person can spend.

I've heard quite the opposite that the fat SPU (or ALU as you've said) doesn't do any DP ops.

While the other SPUs in the US combine together to do DP ops. Therefore you exclude the fat ALUs, and combine the other ALUs, you get the same 1/5 DP FLOPS figure.

Nebojsa not wanting to seem too picky the peak DP rates for the GTX280 are actually 12th of peak SP (or 78 DP GFLOPS for the GTX280 and 90 DP GFLOPS for the higher clocked S1070 cards) because they count the second SP MUL in the peak SP figures but which cannot be dual issued with the DP MAD.

Similarly the peak AMD DP rates are I believe actually 1/5 of peak as it is the fat ALU that does the DP ops and so the peak figures are 200/240/ DP GFLOPS for the 4850/4870 and presumably 240 for the FireStream 9250 when it finally arrives.

No matter the definition the first Teraflop SP single card was the HD3870 X2.

The first single GPU to do it would be the RV770/HD4850

As for the memory scaling and size issue the effect of that won't be fully known until the benefits/limits of the R700's memory/xfire properties are fully divulged. The CrossfireX-Sideport and changes to GDDR5 and the PCI-Express bridge could make it a 2 Teraflop card with similar memory benefits.

The biggest problem for AMD is far fewer options for configuration than TESLA. Rankmount options need ot improve from DAAMIT.

It's far from over this round.

AMD/ATI has the possibility of using "clamshell-mode" for the GDDR5 memories and thereby have 16 mem-chips on a card.
So the 256-bit bus is not as big drawback as you seem to think.
One could argue that Nvidia could use a mem-controller for GDDR5@512-bit to reach double the memory against the competition, but they would need to put 32 mem-chips on the board for that.
Some how I would like to see them try. =)

How do the two, Tesla and Firestream compare internally at the chip level? What about the software? You'll have to wait to read all about that tomorrow. µ

Man u suck .... I was wanting to know more than the title gave....

Yes all this and more can be gathered from the Cineplay 2.0 articles...