I'm sorry to barge in and spoil your party guys, but did you per chance miss the fact that Kepler's shaders will be running at core speed?
Fermi-based cards run their shaders at twice
the core speed.
The new kepler gk104 will have 1536 cuda cores. That is 3 times the gtx 580. The gtx 580 gets 140 mhash/s so wouldn't the new kepler card get 420 mhash/s give or take, or at least 6970 territory?
Unfortunately, I don't think so.
GTX 580's shader clock is 1544 MHz. Let's optimistically assume 1 GHz stock clock speed for Kepler (that translates to 65% of a GTX 580's shader speed).
Therefore, a better estimate might be
3 * 140 * 0,65 * a * b
is overclockability modifier and b
is architecture modifier.
Let's assume that Kepler will overclock to 1200 MHz, what results in a
Since integer operations efficiency has long played second fiddle to floating-point operations, there is no reason to expect huge gains. Let's assume b
= 1.1, i.e. Kepler being 10% better at integer operations than Fermi.
A stock Kepler running its 1536 shaders at 1 GHz would achieve 300 MHash/s
Overclocking the card to 1200 MHz would boost the hash rate to 360 MHash/s
- that's uncomfortably close to a stock VLIW4-based hd6950.
While half of a 7970's hashing speed is far less embarrassing than nVidia's previous generation of GPUs, it's still pretty underwhelming for an expensive and power hungry card a 1536 SP Kelper running at 1 GHz will have to be.