David Kanter (dkanter.delete@this.realworldtech.com) on January 30, 2017 7:49 am wrote:
> juanrga (nospam.delete@this.juanrga.com) on January 29, 2017 4:42 am wrote:
> > David Kanter (dkanter.delete@this.realworldtech.com) on January 20, 2017 2:55 pm wrote:
> > > gallier2 (gallier2.delete@this.gmx.de) on January 20, 2017 7:48 am wrote:
> > > > http://insidehpc.com/2017/01/cray-develop-arm-based-isambard-supercomputer-uk-met-office/
> > >
> > > The original Mont-blanc project had an idiotic abomination, an
> > > Exynos-based supercomputer. AFAICT, it has gone nowhere.
> > >
> > > While Mont Blanc still appears to be pushing ARM, they have
> > > found a much more effective candidate - ThunderX2.
> >
> > The only "idiotic abomination" I can find is the generalized
> > ignorance of some guys about ARM HPC in general
> > and Mont Blanc project in particular. The exynos SoC was selected in a former prototype for exploring
> >

the challenges and benefits of deeply integrated energy-efficient
> > processors and GPU accelerators, compared
> > to traditional homogeneous multicore systems, and heterogeneous CPU + external GPU architectures.

> > If you believed that they would release a final production system based in a Exynos SoC, that
> > is your mistake not their. Those guys know very well that those mobile SoCs aren't suitable
> > for production machines: lack of ECC protection, slow interconnect, low-grade thermal package,
> > 32-bit address space,... "These machines were never intended for production use".
>
> Maybe they should have tried using AMD's Llano, or Intel's client
> SoCs? Both of those would make more sense than an Exynos.

Those aren't ARM SoCs and cannot be used.

> > Moreover your claim "they have found a much more effective candidate" is priceless. Says us, David,
> > how could they explore GPU acceleration on die compared to traditional homogeneous multicore systems
> > using a ThunderX2 SoC that is a homogeneous multicore system and lacks any iGPU?
>
> ThunderX2 is closer to being viable for HPC than a cellphone chip ever will.
> If you want to prototype integrated GPUs only, then just use Intel or AMD.

See above.

> > Similar claims about the current prototype that uses a ThunderX2 SoC. This is not a
> > production system and if "ThunderX2 actually hit targets" or not is rather irrelevant
> > for the goals of the Mont Blanc project. If you bother to read the link that you provide,
> > you will discover that the goals of this third phase of the project are:
> >

> >
• Defining the architecture of an Exascale-class compute node based on the
  > > ARM architecture, and capable of being manufactured at industrial scale;

> >
• Assessing the available options for maximum compute efficiency;
  > >

> >
• Developing the matching software ecosystem to pave the way for market acceptance of ARM solutions.

> >

> > No one with a minimal knowledge of the topic expects an exascale-class system based in 14nm ThunderX2 SoCs.
> >
>
> Actually, ThunderX2 hitting targets does matter. If you want compute efficiency, it's good to have
> a processor that ships on-time, at the target frequency and power consumption. Thunder did none of
> those. It missed frequency by 25%, power by about 30% and shipping by a couple of quarters.

As noted before if ThunderX2 hits targets or not is irrelevant for the goals of the prototype 3.

The miss of targets for the ThunderX chips didn't mean anything either. Mont Blanc used ThunderX in the past to check the software stack.

http://www.prnewswire.com/news-releases/mont-blanc-hpc-software-stack-tested-on-e4s-hpc-platform-based-on-arm64-cavium-thunderx-processor-300112097.html

> > > That makes at least three supercomputers using ARM. Honestly, I'm skeptical of a Cavium-based system,
> > > until we see ThunderX2 actually hit targets. So far Cavium has a bad track record for delivering on their
> > > promises. However, I am very intrigued by Fujitsu's promised Exascale system. They have a good fabric
> > > and understand the system architecture in a way that IBM, Intel, or Nvidia do, and Cavium does not.
> > >
> > > The three ARM systems are all 'national' machines and not bid in a competitive fashion (although
> > > that is true of many HPC systems). I think they will be proving grounds, and if vendors can
> > > show real success there, they may be able to bid for more competitive systems.
> > >
> > > One issue I see is that GPUs tend to require even greater performance from the host CPU to keep
> > > up with Amdahl's Law. That means that a larger number of cores is less attractive compared to
> > > Power9 and Skylake. Also, I expect that server CPUs will start using high-performance DRAM more
> > > commonly. Nvidia would love to see an alternative to Xeon, but so far, they don't have one.
> > >
> >
> > Precisely Nvidia is shipping GPGPUs on ARM-based clusters. I have seen some HPC benches specs in the past
> > and the GPGPUs were running so fast on former ARM-based systems like on Haswell-based Xeons.
>
> For trivial workloads that have little serial code.

Not trivial, but for certain real HPC code.

> > The CPU was
> > not a bottleneck then. Neither it seems the new ThunderX2
> > will be bottleneck. The Chief Technology Director
> > for Extreme Computing at Bull, the company is building the Mont Blanc 3 prototype says:
> >

we expect this to be at the performance level of what you could get with an
> > Intel Skylake Xeon or an AMD “Naples” Opteron.
> >We think that for certain
> > HPC applications ThunderX2 will be at that level – and sometimes, better.

>
> I don't believe this guy.

Your choice.