By: juanrga (nospam.delete@this.juanrga.com), January 7, 2015 5:03 am
Room: Moderated Discussions
Eric Bron (eric.bron.delete@this.zvisuel.privatefortest.com) on January 3, 2015 2:42 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
>
> thank you for the link, he was saying "going away" not "abandoned, which is factually
> true speaking of the near future evolution of MIC, i.e. roughly same number of cores
> in 14 nm KNL than 22 nm KNC, whatever your favorite marketing material says about it
>
> as others have already tried to explain you, it's the opposite to what was "predicted" in a series
> of papers you have posted on this very forum saying that the future is with more simpler cores
"Abandonment" is a synonym of "going away".
My point is that future applications will require the main processor to be a hundred/thousand manycore because multicores don't scale up and cannot provide the needed throughput --the goal is to obtain 50x more performance by about 2020--. I asked Linus twice how he proposes to reach the target goals and he twice avoided the question, he only writes ridiculous rants against manycores.
Reaching this goal will require novel hardware and software techniques now under development. Recently Intel's Crago and coworkers have published a set of latency techniques "to reach the energy efficiency goals of future 1000-core data-parallel processors" [1]. They evaluated and rejected a number of existent latency techniques, including Linus beloved OoO.
Those techniques and others were not available when Intel did start development of KNL. As a consequence KNL engineers had no other choice than a traditional OoO for the KNL core, and then had to increase the SIMD wide to recover part of the lost efficiency.
However, and this is my point, when those novel latency techniques achieve maturity, the big OoO cores don't will be needed and will be replaced by smaller cores. I recall the thread about alternatives to OoO. I gave a selection of interesting papers and the cold response of some people here was "it will not work". I recall Aaron asking me why I believed that mixing VLIW with RA was a good idea. Funny, because we know now that Denver is an in-order VLIW+RA core that outperforms wide OoO cores such as Cyclone or Haswell.
I have diagrams and other technical info about the manycores designed by Nvidia, Fujitsu, IBM, AMD, and others for the 2020, and none use OoO for the throughput cores. KNL is an exception rather than the rule.
NOTES:
[1] http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=6522327&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D6522327
> 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
>
> thank you for the link, he was saying "going away" not "abandoned, which is factually
> true speaking of the near future evolution of MIC, i.e. roughly same number of cores
> in 14 nm KNL than 22 nm KNC, whatever your favorite marketing material says about it
>
> as others have already tried to explain you, it's the opposite to what was "predicted" in a series
> of papers you have posted on this very forum saying that the future is with more simpler cores
"Abandonment" is a synonym of "going away".
My point is that future applications will require the main processor to be a hundred/thousand manycore because multicores don't scale up and cannot provide the needed throughput --the goal is to obtain 50x more performance by about 2020--. I asked Linus twice how he proposes to reach the target goals and he twice avoided the question, he only writes ridiculous rants against manycores.
Reaching this goal will require novel hardware and software techniques now under development. Recently Intel's Crago and coworkers have published a set of latency techniques "to reach the energy efficiency goals of future 1000-core data-parallel processors" [1]. They evaluated and rejected a number of existent latency techniques, including Linus beloved OoO.
Those techniques and others were not available when Intel did start development of KNL. As a consequence KNL engineers had no other choice than a traditional OoO for the KNL core, and then had to increase the SIMD wide to recover part of the lost efficiency.
However, and this is my point, when those novel latency techniques achieve maturity, the big OoO cores don't will be needed and will be replaced by smaller cores. I recall the thread about alternatives to OoO. I gave a selection of interesting papers and the cold response of some people here was "it will not work". I recall Aaron asking me why I believed that mixing VLIW with RA was a good idea. Funny, because we know now that Denver is an in-order VLIW+RA core that outperforms wide OoO cores such as Cyclone or Haswell.
I have diagrams and other technical info about the manycores designed by Nvidia, Fujitsu, IBM, AMD, and others for the 2020, and none use OoO for the throughput cores. KNL is an exception rather than the rule.
NOTES:
[1] http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=6522327&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D6522327