Kepler Nvidia GeForce GTX 780

[Lanek]

well, i see another problem there, i think thoses spec are fake somewhere, but a quick calculation.. 12geometry engine, 96TMU said 12 SM x 48SP = 576SP ( the 560 = 8x48SP ) the 580 had 32SP x 16SM .

560 = half the number of SM you find in the 580. ( the easiest way for scale the architecture and have scaling for size, cost,TMU, Geometry in the line up )

GK110 could well be: 24SMx32SP = 768SP + 24 polymorph engine. ( 2x the number of SM you find in the GK104 ). if thoses spec are right, we are back to initial speculation.

[***Deimos***]

Just look back who was right last time with Fermi launch:

FACT

AN00BIS was right!
- almost 6months before GTX480 launch
http://www.xtremesystems.org/forums/...ia-gt300/page4


"Pessimists" like me got some right
- near 300W
- huge chip with fab issues
- clocks dont scale to 5870 levels.
- at least 20% faster than 5870.

Wrong:
- DX11 Tessellation and Compute driver support lot better than expected - Home Run hit!
- nowhere near expected $699
- Said it was not feasable, but they made a 2xFermi = GTX 590.

========================

Kepler Prediction
- a bit smaller than previous high end - 490mm2
- 5.5B trans... likely 960 shader /ALU 24x40?? ... (tiny chance 1124 shader... designer probably worried about yields)
- guessing mainstream parts will be like 640SP.
- much higher GPU clocks - unlike AMD sheer number of units, much higher TLP, and ILP.
- doubt anybody's willing to risk moving away from 5.5Ghz+ GDDR5
- nVidia has a lot of experience with 384bit memory - one more time!
- its a given, the shader organization will be different, and more fine grained (ie 4 sections each with their own L2, tesselation etc).


mainstream SP trend is: 128 >88%> 240 >113%> 512....960
112 >>50%>>192>>75%>>336 .... 640-720