Knights Landing CPU Speculation

Article: Knights Landing CPU Speculation
By: Amiba Gelos (b98202036.delete@this.ntu.edu.tw), November 21, 2013 5:14 pm
Room: Moderated Discussions
Linus Torvalds (torvalds.delete@this.linux-foundation.org) on November 21, 2013 3:12 pm wrote:
> Amiba Gelos (b98202036.delete@this.ntu.edu.tw) on November 19, 2013 8:36 pm wrote:
> >
> >
> > Second, I wonder whether the claim that Knights Landing is bootable really implies we need good scalar
> > core. A major advantage of bootable accelerator is that by
> > moving from a heterogeneous MP to SMP the programming
> > paradigm is greatly simplified. Therefore, faster scalar execution is not really needed.
>
> People definitely need faster scalar execution.
>
> Guys, don't look at just DGEMM benchmarks or pure "throughput" stuff. The real HPC market
> cares deeply about things like "sufficient NFS client performance" for example, because in
> reality, that's how most of those clusters end up doing a lot of the data distribution.
>
> And the old Atom core was too weak to do even that. Seriously. I know, because I had people complain
> to me about it, and talk about trying to do special tricks just to get NFS performance up.
>
> (Btw, don't get me wrong. It's not always NFS, but it almost always is something like
> NFS. Sometimes it's fancier specialized filesystems, but that doesn't really change the
> basic argument: there's a very noticeable component that isn't about computation).
>
> So scalar performance absolutely does matter. Not necessarily for benchmarks, where
> access patterns are so simple that you can feed the data using very specialized channels,
> or there is no "real" data to begin with. But people who buy those monster machines
> based on peak numbers end up finding that they really do need other things too.
>
> There are almost no actual real HPC loads that only want floating point, particularly the GPU kind of floating
> point with special private memory. Pretty much all the real ones have big datasets that they are trying to work
> with, and it's not even just "communication" between nodes, it's really more fundamental than that.
>
> Prediction: the whole PCIe card model will die, very much due to this kind of pressure.
> Sure, you can get monster peak flops with the GPU approach and using DMA-to-host-memory,
> but it's painful as hell for big datasets. Because much of the real world ends up wanting
> more than just a floating point accelerator with an inconvenient memory interface.
>
> Linus

Well, I'm not really a HPC guy, but hell sure the filesystem is a neat problem, and these kinds of problems requires blazing fast cores.
Heard rumors that guys at CERN have some solutions though, but I guess they have to rewrite their ancient analysis code first :-)

As far as I know, stuffs like numerical GR, Quantum MC and lattice QCD all have very nice access pattern and relatively small dataset yet require large amount of data transfer and raw flops.
Since most of the scientific HPC programs are either PDE or Monte Carlo problems I can't really see why we cannot do prefetch manually (which was done manually).

BTW, I totally agree w/ you that private memory is painful :D
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
Knights Landing CPU SpeculationDavid Kanter2013/11/18 02:03 AM
  Knights Landing CPU Speculationnone2013/11/18 02:59 AM
    Knights Landing CPU SpeculationPatrick Chase2013/11/23 03:18 PM
      Knights Landing CPU SpeculationAdam Kalisz2013/11/26 01:20 AM
        Over 2,000 mm^2 of eDRAM???Mark Roulo2013/11/26 09:28 AM
          Over 2,000 mm^2 of eDRAM???David Kanter2013/11/26 11:09 AM
          Over 2,000 mm^2 of eDRAM???Eric Bron2013/11/26 11:21 AM
            Over 2,000 mm^2 of eDRAM???tarlinian2013/11/26 11:50 AM
              Over 2,000 mm^2 of eDRAM???Eric Bron2013/11/26 01:07 PM
                Over 2,000 mm^2 of eDRAM???Eric Bron2013/11/26 01:09 PM
                Over 2,000 mm^2 of eDRAM???aaron spink2013/11/26 03:03 PM
                  Over 2,000 mm^2 of eDRAM???Eric Bron2013/11/26 11:42 PM
                    Over 2,000 mm^2 of eDRAM???aaron spink2013/11/27 10:31 AM
              Over 2,000 mm^2 of eDRAM???David Kanter2013/11/26 04:25 PM
                Over 2,000 mm^2 of eDRAM???tarlinian2013/11/26 07:01 PM
          Over 2,000 mm^2 of eDRAM???Eric2013/11/27 02:54 AM
            eDRAM is DRAM in a logic-oriented processPaul A. Clayton2013/11/27 07:10 AM
  Knights Landing CPU SpeculationJames2013/11/18 05:26 AM
    Knights Landing CPU SpeculationMichael S2013/11/18 02:57 PM
      Knights Landing CPU SpeculationUrban Novak2013/11/19 12:49 AM
        Knights Landing CPU Speculationnone2013/11/19 01:19 AM
          Knights Landing CPU SpeculationEric2013/11/19 07:48 PM
            Total GPGPU/Xeon Phi market maybe ~ $500M/year ...Mark Roulo2013/11/20 10:35 AM
      Knights Landing CPU SpeculationWes Felter2013/11/19 12:06 PM
        Knights Landing CPU SpeculationMichael S2013/11/19 12:49 PM
  Knights Landing CPU SpeculationEric2013/11/18 12:17 PM
    Knights Landing CPU SpeculationDaniel2013/11/19 02:28 AM
      Knights Landing CPU SpeculationEric2013/11/19 07:36 PM
      HPC guys score FLOPS non-obviouslyMark Roulo2013/11/20 10:43 AM
        3-TFlops-DGEMMMichael S2013/11/20 10:59 AM
          3-TFlops-DGEMMMark Roulo2013/11/20 12:22 PM
            3-TFlops-DGEMMDaniel2013/11/20 01:04 PM
              3-TFlops-DGEMMEric2013/11/21 01:28 AM
                3-TFlops-DGEMMMichael S2013/11/21 05:48 AM
                  3-TFlops-DGEMMRecessionCone2013/11/21 11:13 AM
                    3-TFlops-DGEMMMichael S2013/11/21 02:34 PM
                  3-TFlops-DGEMMEric2013/11/22 02:10 AM
                    3-TFlops-DGEMMMichael S2013/11/22 04:41 AM
                    A (not very sensible) alternative: FMADD + FADDPaul A. Clayton2013/11/22 08:19 AM
                  3-TFlops-DGEMMSylvain Collange2013/11/24 02:37 AM
                    3-TFlops-DGEMMMichael S2013/11/24 06:06 AM
                      3-TFlops-DGEMMSylvain Collange2013/11/24 09:28 AM
        HPC guys score FLOPS non-obviouslyPatrick Chase2013/11/23 02:58 PM
  Knights Landing CPU SpeculationPaul Caheny2013/11/18 01:25 PM
    Knights Landing CPU SpeculationKonrad Schwarz2013/11/19 12:24 AM
  Knights Landing CPU SpeculationAmiba Gelos2013/11/19 07:36 PM
    Knights Landing CPU SpeculationDavid Kanter2013/11/20 09:52 AM
    Knights Landing CPU SpeculationLinus Torvalds2013/11/21 02:12 PM
      Knights Landing CPU SpeculationAmiba Gelos2013/11/21 05:14 PM
      Knights Landing CPU SpeculationPatrick Chase2013/11/23 03:33 PM
        Knights Landing CPU SpeculationLinus Torvalds2013/11/25 11:29 AM
          Knights Landing CPU SpeculationLinus Torvalds2013/11/25 12:05 PM
          Knights Landing CPU SpeculationPatrick Chase2013/11/25 12:22 PM
            Knights Landing CPU SpeculationLinus Torvalds2013/11/26 10:11 AM
          Knights Landing CPU SpeculationEric2013/11/26 03:05 AM
            Knights Landing CPU SpeculationEric2013/11/26 03:15 AM
            Knights Landing CPU Speculationnone2013/11/26 03:33 AM
              Knights Landing CPU SpeculationEric2013/11/26 06:30 PM
                Knights Landing CPU SpeculationEric2013/11/26 06:34 PM
                  What is MCDRAM?anon2013/11/26 08:58 PM
                    What is MCDRAM?none2013/11/27 01:00 AM
                      What is MCDRAM?Klimax2013/11/27 02:19 AM
                  Knights Landing CPU SpeculationKlimax2013/11/26 11:06 PM
                Knights Landing CPU SpeculationKlimax2013/11/26 11:05 PM
            Knights Landing CPU Speculationanon2013/11/26 05:53 AM
              Knights Landing CPU Speculationnone2013/11/26 06:20 AM
                Knights Landing CPU SpeculationMichael S2013/11/26 08:06 AM
                  Knights Landing CPU Speculationnone2013/11/26 09:18 AM
                    Knights Landing CPU SpeculationEric Bron2013/11/26 01:21 PM
                      Knights Landing CPU SpeculationEric Bron2013/11/26 01:27 PM
                        Knights Landing CPU Speculationnone2013/11/26 02:26 PM
                    Knights Landing CPU Speculationanon2013/11/26 05:42 PM
                      Knights Landing CPU Speculationnone2013/11/27 01:08 AM
                        Knights Landing CPU Speculationanon2013/11/27 01:50 AM
                          Knights Landing CPU Speculationnone2013/11/27 01:58 AM
                      Knights Landing CPU SpeculationMichael S2013/11/27 01:25 AM
                        Knights Landing CPU Speculationanon2013/11/27 02:32 AM
                          Knights Landing CPU SpeculationMichael S2013/11/27 03:08 AM
    Knights Landing CPU SpeculationChung Leong2013/11/27 01:28 AM
      Knights Landing CPU SpeculationMichael S2013/11/27 02:53 AM
        Knights Landing CPU SpeculationChung Leong2013/11/27 01:03 PM
  BiG.LiTTLe for KNL?Jeff K2013/11/22 06:17 AM
    BiG.LiTTLe for KNL?Patrick Chase2013/11/23 02:54 PM
      BiG.LiTTLe for KNL?Patrick Chase2013/11/23 03:01 PM
  Transactional memoryPatrick Chase2013/11/23 02:37 PM
    Transactional memoryBhima2013/11/25 07:01 AM
      Transactional memoryPatrick Chase2013/11/25 11:52 AM
  Knights Landing CPU SpeculationDaniel2013/11/25 02:17 AM
    Knights Landing CPU SpeculationKlimax2013/11/25 03:12 AM
    Knights Landing CPU Speculationnone2013/11/25 04:05 AM
      Knights Landing CPU SpeculationKlimax2013/11/25 04:45 AM
        Knights Landing CPU Speculationnone2013/11/25 04:55 AM
          Knights Landing CPU Speculationgmb2013/11/25 07:21 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell green?