Here -- I'm confused what point 1 is referring to in terms of execution units per core, as doesn't execution units per core concern SIMD (point 2)? Or is this referring to superscalar parallelism? So point 1 means something like: "have at least n threads of work (where n = num cores), and have sufficient independent operations in each such that we can maximally utilize superscalar abilities"?

having sufficient parallel work to utilize all available execution units seems to fit certain problem domains better that others. If your problem can't be parallelized, are you out of luck?

@apappu, I think slide 79 might help clarify what @kayvon means by "model parallel processors". Essentially, there is a mix of multiple cores running independent instructions amongst themselves, and these instructions are SIMD. In the example on slide 79, he had 16 cores, each with 4 threads (4 execution contexts / instructions stream), with 8 ALU's per core (8-wide SIMD capability) = 512 independent pieces of work needed to utilize all available execution units.

In other words, I think "modern processors" imply SIMD processing.