By: juanrga (nospam.delete@this.juanrga.com), August 26, 2015 11:11 am
Room: Moderated Discussions
Aaron Spink (aaronspink.delete@this.notearthlink.net) on August 26, 2015 7:25 am wrote:
> juanrga (nospam.delete@this.juanrga.com) on August 25, 2015 6:01 pm wrote:
> >
> > Nvidia talk at ISC2015 was much more interesting. They compared
> > two KNL CPUs against Power+CUDA using Amdahl's
> > law. At 98% parallel the KNL was competitive. At 90% parallel
> > work the KNL system was about two times slower than
> > the Power+CUDA system: ~2 min vs 4.5 min. At 70% parallel work, the KNL system was more than 3x slower.
> >
> > Wider vector units and less cores had worked better.
> >
>
> So much bogusity... First, that's not what he did. What he did was try to play up the low "node" count on
> summit. They get this node count by disregarding that their "node" is actual several nodes. Second, he then
> tried to apply some voodoo computer science by saying higher node count equals lower performance (which funny
> enough has actually been *disproven* by their partner IBM in actual practice!). If you want to do an actual
> Amdahl's law evaluation, you have to compare the total parallelism requirements of both systems, which are
> both incredibly high. If anything, it is likely that the Intel system will be running at a higher frequency,
> thus requiring less parallelism for a given level of performance, but since no one really outside of a select
> few in the actual government know the actual *paper* specifications of the systems...
>
There wasn't voodoo at the talk. He simply explained why an heterogeneous system (what Nvidia calls LOC+TOC approach) offers superior performance and efficiency to homogeneous systems. And since he compared Summit to Aurora he compared Power-CUDA to Intel MIC.
> juanrga (nospam.delete@this.juanrga.com) on August 25, 2015 6:01 pm wrote:
> >
> > Nvidia talk at ISC2015 was much more interesting. They compared
> > two KNL CPUs against Power+CUDA using Amdahl's
> > law. At 98% parallel the KNL was competitive. At 90% parallel
> > work the KNL system was about two times slower than
> > the Power+CUDA system: ~2 min vs 4.5 min. At 70% parallel work, the KNL system was more than 3x slower.
> >
> > Wider vector units and less cores had worked better.
> >
>
> So much bogusity... First, that's not what he did. What he did was try to play up the low "node" count on
> summit. They get this node count by disregarding that their "node" is actual several nodes. Second, he then
> tried to apply some voodoo computer science by saying higher node count equals lower performance (which funny
> enough has actually been *disproven* by their partner IBM in actual practice!). If you want to do an actual
> Amdahl's law evaluation, you have to compare the total parallelism requirements of both systems, which are
> both incredibly high. If anything, it is likely that the Intel system will be running at a higher frequency,
> thus requiring less parallelism for a given level of performance, but since no one really outside of a select
> few in the actual government know the actual *paper* specifications of the systems...
>
There wasn't voodoo at the talk. He simply explained why an heterogeneous system (what Nvidia calls LOC+TOC approach) offers superior performance and efficiency to homogeneous systems. And since he compared Summit to Aurora he compared Power-CUDA to Intel MIC.