Originally Posted by SlowPWNY
what if by "trimming the fat" of commercial features of the chip they could get all 15 smx's to work? 1536 shaders to 2888 is like a 84% increase and then of course the additional mem bandwidth. if by some miracle they could get all the SMX's pumping and sold the card it would explain why it would be a "limited" run.
BRAIN FART^^^^ i was comparing it to a single 680
lets say by the grace of the silicon gods they get all 2888 shaders working. they would only need to clock it in around 830mhz to get within the 85% mark of a 690 if i did my math right.
It is proven that Kepler is mostly limited by it's ROP/L2/memory-parts, and that additional SMXes doesn't add much to actual performance. Considering most Kepler cards scale pretty well with core clock increases, the most likely assumption is that it's the ROPs that are the main bottleneck
The GK110 most likely has 48 ROPs (Nvidia hasn't disclosed the amount,) which is 50% more than the GTX 680 (25% less than the 690,) and if this Titan is supposed to have 85% of the 690s performance it will need to run at 1037+ MHz core. A GK110 running at 1037+ MHz core would consume far too much power to be acceptable for anyone other than rabid fanboys and K1NGP1N
Bandwidth-wise, there's no way the Titan will have more bandwidth than the 690 (Titan would need to run it's memory at 2000MHz to deliver identical bandwidth,) GDDR5 is not yet available at 2000MHz and getting a 384-bit memory controller to scale that far up in clock speeds when the 256-bit bus of the GK104 caps out arund 1800MHz is wishful thinking at best.
The math simply doesn't add up to this Titan having "85% of the 690s performance," the GK110 simply isn't powerful enough for such a thing to happen.