By: Aaron Spink (aaronspink.delete@this.notearthlink.net), January 10, 2015 9:03 pm
Room: Moderated Discussions
juanrga (nospam.delete@this.juanrga.com) on January 10, 2015 8:21 pm wrote:
> The class is very rich in its diversity, but this is not different to everything
> else: there are different kind of CPUs, of cars, of people...
>
No it isn't rich in diversity, its rich in papers. The actual implementations are all pretty much the same.
> Manycores are not defined by number of cores, but by the kind of compute optimization.
> The higher core count in GPUs and Phi compared to multicores like Haswell or Power8
> is a result, a byproduct, of that microarchitectural optimization.
>
Once again, GPUs are actually low core count devices. Don't confuse marketing and ALU lanes with cores. The top end Haswell parts have more cores than most GPUs. And as far as uArch optimization, there is actually very very little of that in Phi. Its using pre-existing off the shelf Intel cores and attaching a wide SIMD engine. None of the Phi products have used a from the ground up core and they've all made minimal changes to those pre-existing cores. Of which it should be pointed out, none of the Phi cores have been designed for manycore. Unless you want to consider P5 a manycore design or silvermont a manycore design, in complete disregard of all evidence and reality.
And the reality is that modern GPUs have nothing to do with manycores. They are all based off of deeply threaded wide SIMD processors. They have more in common with old school barrel threaded vector machines than they do with manycore.
> The class is very rich in its diversity, but this is not different to everything
> else: there are different kind of CPUs, of cars, of people...
>
No it isn't rich in diversity, its rich in papers. The actual implementations are all pretty much the same.
> Manycores are not defined by number of cores, but by the kind of compute optimization.
> The higher core count in GPUs and Phi compared to multicores like Haswell or Power8
> is a result, a byproduct, of that microarchitectural optimization.
>
Once again, GPUs are actually low core count devices. Don't confuse marketing and ALU lanes with cores. The top end Haswell parts have more cores than most GPUs. And as far as uArch optimization, there is actually very very little of that in Phi. Its using pre-existing off the shelf Intel cores and attaching a wide SIMD engine. None of the Phi products have used a from the ground up core and they've all made minimal changes to those pre-existing cores. Of which it should be pointed out, none of the Phi cores have been designed for manycore. Unless you want to consider P5 a manycore design or silvermont a manycore design, in complete disregard of all evidence and reality.
And the reality is that modern GPUs have nothing to do with manycores. They are all based off of deeply threaded wide SIMD processors. They have more in common with old school barrel threaded vector machines than they do with manycore.