systemfehler
Member
Correct. Its a balancing act. Instead of wasting GPU resources trying to make the machine juggle expensive compute tasks, a portion of the GPU is shelved off to make these tasks relatively cheap. It is a better solution. Hence the 1.4 TF left over will be very efficient at what it does.
A VLIW architecture already depends that the compiler knows whats best by grouping tasks accordingly. I would rather have seen 18 CUs for GPGPU as I fear that the Jaguar cores might need all the help they can get.