Linux And Not-Xeon Take The Cake

See Linux Now Runs on All of the Top 500 SupercomputersIn an article about supercomputers and Linux now running rampant over them, I found this nugget:

Of the top two fastest machines. One’s a Xeon and the other’s a Sunway. The Xeon gets 3 Tflops per kW and the Sunway gets 8. There’s a good reason to avoid Intel products, IMHO.

In fact, the Chinese machine is not the “greenest” machine in the list. The Green500 list shows several Xeon machines do better but they are not #1 in performance… so it’s a competitive matter and Intel has no monopoly on performance these days.

About Robert Pogson

I am a retired teacher in Canada. I taught in the subject areas where I have worked for almost forty years: maths, physics, chemistry and computers. I love hunting, fishing, picking berries and mushrooms, too.
This entry was posted in technology and tagged , , , , , . Bookmark the permalink.

7 Responses to Linux And Not-Xeon Take The Cake

  1. Deaf Spy says:

    Your statement makes no sense.

    It does in the context of Robert’s thoughts on CPUs. It’s been a looong lasting fun with him.

  2. Dylan Hunt wrote, “Ever wonder why DPRK does live tests, while Russia, China and the US continue to modernize their nuclear arsenals without having to live test? They can afford such toys, DPRK cannot.”

    DPRK has gone nuclear on a low tech path. They don’t really care about the cost of testing nukes but just that they get nukes. A simulator is just about useless to anyone who does not have the data collected from many tests of nukes. If DPRK wanted a super-computer for doing simulations they will get them. Otherwise, they’ll do it the old fashioned way. Remember, the original nukes were developed with slide rules, lots of pencils and chalk and one computer using vaccuum tubes. Deaths attributed to the Manhattan Project were of lab geeks. Louis Slotin, a Canadian physicist from the physics department of my alma mater, was doing a test in a lab to determine the critical mass for a nuke. He went a bit too far but managed to prevent a melt-down while irradiating himself to death. DPRK is way more advanced than the Manhattan Project but they still can get things done without simulations.

  3. Dylan Hunt says:

    Let’s compare a specialized chip against a multipurpose chip that can run x86 code better than anything else, and declare victory.

    Your statement makes no sense. The SW26010 (used in the Sunway TaihuLight) is a general purpose RISC design based on DEC’s Alpha, scaled up to 260 cores. It isn’t special purpose.

    We’re talking HPC and scientific computing*. Fortran and Python are king, and nobody cares about legacy binary support. That Xeons run code compiled specifically for x86 better than anything else is of no consequence. SW26010 runs code compiled for Alpha better than anything else, what’s your point?

    Robert on the other hand is misguided on performance, as well as the background regarding why the SW26010 was developed.

    TaihuLight (SW26010) has more than 3x as many cores as Tianhe-2 (Xeon & Xeon Phi), for 3x the sustained performance and just over twice the theoretical peak performance, while using slightly less power. Tianhe-2 delivers 17 megaFLOPs (0.017 TFLOPs) per core vis-avis Sunway’s 11 megaFLOPs (0.011 TFLOPS) per core. It’s really not a function of how well it runs the x86 compiled code, we’re talking

    The fact of the matter is that Tianhe-2 would slightly edge out TaihuLight in terms of raw performance if it tripled up on the core count (so it utilizes as many cores as TaihuLight), but would draw just under 3x as much power.

    For a 100 TFLOP cluster, SW26010 beats out the Xeons, mainly on power draw. They could have gone with the UltraSparc VIIIfx in Riken’s K computer (TaihiLight has 15 the cores, for 8x the performance, while operating at 64% efficiency vs Riken’s 93% efficiency*) Given the per-core output and efficiency considerations, it would take far, far fewer just over 1/10th the SparcFX cores to build a cluster with the same sustained TFLOPs output as TaihuLight.

    But none of this was a consideration for the decision to design and implement SW26010.

    TaihuLight (as well as Tianhe-2) were commissioned by the Chinese military. Obama instituted a ban on the sale of Xeon and Nvidia tech to the Chinese government. It’s unlikely that the Japanese government would allow Fujitsu to sell the Chinese military SparcFX as well.

    These things are used to run the simulations necessary for China to pursue the development of hypersonic delivery systems and the modernization of their nuclear arsenal while upholding their treaty obligations (Beijing has signed treaties barring live testing). Beijing didn’t go with Sunway as a function of performance or greenness, it comissioned the development of the SW26010 because it had no choice but to roll its own high performance architecture.

    Ever wonder why DPRK does live tests, while Russia, China and the US continue to modernize their nuclear arsenals without having to live test? They can afford such toys, DPRK cannot.

  4. oiaohm says:

    Let’s compare a specialized chip against a multipurpose chip that can run x86 code better than anything else, and declare victory.
    That we don’t know Deaf Spy.
    https://venturebeat.com/2015/09/03/chinas-loongson-makes-a-64-bit-mips-processor-that-can-run-x86-and-arm-code/
    Mips processors made and designed in china have extensions that make x86 emulation perform way better. So china based mips super computer and xeon based supercomputers both can run x86 code we just don’t have a benchmark in pure x86 to work out if xeon is that competitive any more.

    Of course while only used inside China Intel cannot use their x86 patent collection against these mips processors.

    Deaf Spy basically this is you again claiming something you don’t have evidence for. Xeon might run x86 better than everything else. But there is enough different processes these days with extensions to accelerate x86 emulation that we no longer can be sure of that. Of course we living in areas where Intel can use their x86 patent pool are not seeing any of the competing chips.

  5. Deaf Spy says:

    Let’s compare a specialized chip against a multipurpose chip that can run x86 code better than anything else, and declare victory.

    Pathetic.

  6. Grece wrote, “To date, Bob has not purchase any hardware to upgrade Beast.”

    The first incarnation of Beast cost about $1500 and used a single 32-bit core. The next version was AMD64. The third incarnation uses quad-core. ISTR it used Debian mostly. The first one might have used RedHat/K12LTSP. Storage has ramped up. Networking has ramped up. Beast just happens to be big and powerful enough for what I do. There isn’t much need for an upgrade except to reduce power-consumption. That will happen sooner rather than later. It would have happened already if the Cello had come through.

  7. Grece says:

    But not a one iota of Linux desktops in the wild.

    Oh avoid Intel at your own peril!

    Be like Bob, attempt to buy ARM boards and string a terminal server with cheap thin clients, then boast about it for months, with nothing to show for it. To date, Bob has not purchase any hardware to upgrade Beast.

Leave a Reply