EduardoS (no@spam.com) on 3/23/12 wrote:
---------------------------
>anon (no@thanks.com) on 3/23/12 wrote:
>---------------------------
>>For instance, the
>>techreport review claims that there are 4 schedulers inside an SMX, each tied to
>>a distinct set of 48 execution units.
>
>If tying untis at 2 schedullers they could have 3 32 width units per two schedullers.
>
>Why making things more complex? They could just double the width of the SIMD units
>but inestead they put half the execution units at 32 work-items and the other half at 16?
>
I'm both utterly surprised and disappointed, by the lack of concrete evidence or any official information on the SIMD configuration in Kepler. As it goes, all we have is pure FUD and contradicting guess work.
On the other hand, this is to be expected (sort of), as GK104 is intended primary as a "gaming" architecture and obviously Nvidia doesn't feel the need to expose/market the compute aspect details in a hurry.