No, I need moar corez!!
Well I'm assuming that in 1-2 years more software will be using all cores it can see.
I'm not expecting too much with these instruction extensions, they are mostly limited to specific situations. AVX helps SB a lot in a few cases, but hardly a game changer compared to predecessor. I could be wrong, but I'm skeptical that AVX2 will bring much more.
With all due respect, you are dead wrong. ^_^
Quiz question: how many cores does a modern GPU have? If you think more than one thousand; you've been fooled by marketing. The thing is, they count each SIMD lane as a separate core. Using the same logic, mainstream Sandy Bridge chips would already have
64 cores (running at 3-4 times the frequency).
Next, AVX2 is a far bigger deal than AVX. For the first time in the history of mainstream processors, every (relevant) scalar instruction will have an SIMD vector equivalent! This is fundamental because it makes it possible to vectorize most performance-critical code loops.
Note also that loop vectorization is an optimization that can be handled automatically by compilers, and will be supported by all major ones
before Haswell hits the street. Furthermore, throughput computing frameworks like OpenCL can seamlessly support GPUs and CPUs (with or without AVX2) so application developers don't even have to explicitly target AVX2 to get a speedup.
And last but not least, more cores are worthless if the synchronization between them is slowing things down. There's a
quadratic relation for the number of possible synchronization interactions. Two cores have only one iteraction, four cores have six, eight cores have twenty-eight! Haswell improves the synchronization with
hardware transactional memory. So you're getting better performance scaling, even with the same number of cores.
Ivy Bridge is a minor evolutionary step, while Haswell is without a doubt revolutionary.