juanrga (nospam.delete@this.juanrga.com) on January 7, 2015 5:47 am wrote:
> Michael S (already5chosen.delete@this.yahoo.com) on January 3, 2015 1:30 pm wrote:
> > juanrga (nospam.delete@this.juanrga.com) on January 3, 2015 1:11 pm wrote:
> > > Michael S (already5chosen.delete@this.yahoo.com) on January 3, 2015 12:36 pm wrote:
> > > > juanrga (nospam.delete@this.juanrga.com) on January 3, 2015 12:02 pm wrote:
> > > > > Eric Bron nli (eric.bron.delete@this.zvisuel.com) on January 2, 2015 2:28 pm wrote:
> > > > > > > I still recall when he pretended that Intel had abandoned manycores with the new Xeon
> > > > > > > Phi
> > > > > >
> > > > > > link ?
> > > > >
> > > > > http://www.realworldtech.com/forum/?threadid=143796&curpostid=143824
> > > >
> > > > So, Linus said that KNL is multicore called manycore fore marketing purpose. Where is he wrong?
> > > >
> > >
> > > On that KNL is a manycore not a multicore.
> >
> > KNL cores are relatively big and fully cache-coherent to each other.
> > To me it sounds much more like multicore than manycore.
> > The only manycore-like feature is the absence of on-chip LLC.
>
> Big and small are relative terms, but KNL cores are smaller than Skylake cores.
>
> Cache-(in)coherence and on-chip LLC don't characterize the difference between many- and multicores.
>
> > Anyway, in the absence of consensus definitions of manycore
> > and multicore we can argue about it ad infinitum.
>
> The differences between a Nvidia GPGPU and a Power8 CPU

And then you replace Power8 in your statement by Oracle SPARC-M7. Is difference still obvious?

> or between an Xeon
> Phi and a Skylake Xeon are real, we don't need any consensus to get that.
>

Indeed, ultra-high-end NVidea GPGPUs do have few dozens of blocks that could be reasonably called "cores" (I consider Maxwell SMM to be equivalent of quad-core CPU module). And indeed few dozens *are* many. So I agree to call them manycore. However I don't think that you'll find many people like me that would agree to call them "manycore".

As to KNL and Linus, he said that KNL is "more multicore"="less manycore" than it's direct predecessor KNC. It's hard to argue against it.

But I'd guess that was not even his prime point. His prime point was that neither KNC nor KNL represents any sort of "future" for general-purpose computers. KNL, may be, makes less bad general-purpose computer than KNC, but on absolute scale of usefulness for anything except a narrow number-crunching niche, even KNL is horrible.