in other words, I suggest distinguishing between simd flops and mimd flops in that the latter only counts operations happening in parallell if those operations are randomly distinct. i.e. it's the maximum flops you'll get if all your operations are impossible to run on simd. in pretty much all cases it's just min(number of instruction-queues , number of float operating queues) times frequency of operation-execution. for example amd's processor has 4 cores (at 4ghz) and enough ressources to perform their calculations. the gpu part of apu has 32 wavefronts at 1ghz each. a total of 48 Gflops mimd. now if there was a processor with madd taking 4 parameters, to effectively perform add and multiply for 2 distinct pairs of variables in parallell, and a possibility to perform 2 mul and 2 add instructions in a simd way, then this processor may double its mimd flops-count. same with intel-processors and their hyperthreading: there should be no bottle-neck if the flops happening in parallell are chosen badly...
needless to say I have no cs education at all, above words are just my half-informed opinion. I'm not even a hobbyist in computer science.Statistics: Posted by piotr5 — Sat Oct 24, 2015 7:02 am
]]>