Sheevaplug vs Efika MX vs Nvidia Tegra2 vs TI OMAP4 Pandaboard: benchmarks

Hi everyone,

I’m writing this post just to publish some results of benchmarks i’ve done lately on the ARM devices i have. They seem pretty popular, so i guess i could publish them on my blog so other people have it easy to find them.

== memcpy ==

The most popular one is the memcpy benchmark. It was written by a gentoo user and friend Siarhei Siamashka (ssvb is his nick on He published it on:
I simply ran it on the devices i have.

You can get the latest results of this benchmark on the devices i’ve been able to run the benchmark on the following page:

== Gentoo build times ==

If you’ve read my blog before, you’ll know i have a page where i list the buildtimes of different basic packages(binutils, gcc, glibc and portage) on different kind of machines to compare it. The page is here:

Unfortunately those results have a problem, i only did a part of them myself, the rest were provided by other people, and as you can see they were using different kind of storage… so last year i decided i could do a new page where all the build times were using the same storage (USB stick), same package versions and same configuration…

The result is this one:

Hope you find them useful.


18 Responses to “Sheevaplug vs Efika MX vs Nvidia Tegra2 vs TI OMAP4 Pandaboard: benchmarks”

  1. Matija "hook" Šuklje Says:

    I’ve searched high and low and I still haven’t been able to see how powerful SheevaPlug etc. are in practice.

    What I’m aiming for is using it as my personal server and run on it Drupal, GNUsocial, ownCloud, RoundCube, Postfix, an XMPP server, maybe YaCy… would it be able to cope with that?

    Sorry for being slightly off-topic, but you’re the only person I know of that runs Gentoo on it :\

    • Raúl Porcel Says:

      Well, it depends on what kind of stress do those applications produce to the machine. I’d say the sheevaplug can cope with what you’re asking…have you tried asking in the sheevaplug forums?

      • Matija "hook" Šuklje Says:

        Not yet. But I asked on irc:// and some admin there said any dynamic content (PHP, Python, Perl, Ruby etc.) wouldn’t be a smart idea. And that double goes for Java (e.g. YaCy).

        BTW, is Gentoo a good solution for a *Plug or is a binary distro better suited?

      • Raúl Porcel Says:

        Hi Matija,

        sorry for the late reply.

        Well, i can tell your for sure that at gentoo we don’t yet support java on arm. Simply because no one of the devs has any interest on it. Definitely java stuff is going to work better on a more common architecture. Obviously an x86 box is more suited for web pages, it highly depends on how much traffic do you have and whats the performance you expect from it. Keep in mind that an x86 machine will always(for some years, at least) beat an arm machine in all aspects.

        Regarding if gentoo is a good solution. Well, depends, compiling stuff, like gcc, on the plug takes around 5 hours. That would mean having the cpu at 100% of usage during 5 hours. And i always recommend that before installing gentoo on a device like this, you should know how gentoo works(ie, having another machine with gentoo).

        Hope it helps

  2. zou Says:

    While these tests don’t represent a lot it still shows the tegra2 and the omap4 are pretty similar (graphics and DSP excluded) (well, they should be anyway)

    it would be interesting to run a bench suite on them, including 3D, video encode, etc

  3. zou Says:

    Sorry for the double post, actually i wonder if you would be able to run some real encode such as a small 720p source on x264 software+neon (no proprietary DSP) on both tegra and omap boards and tell us the difference in FPS.

  4. Karel Gardas Says:

    Raul, could you also be so kind and run “standardized” stream benchmark on your available hardware?
    Thanks! Karel

  5. Matt Turner Says:

    It should be noted that Raúl’s Pandaboard has slower memory than what is used on standard Pandaboards.

    • Siarhei Siamashka Says:

      > It should be noted that Raúl’s Pandaboard has slower memory than what is used on standard Pandaboards.

      It should be also noted that Raúl’s Pandaboard does not have cache controller properly configured (its kernel is missing the following patch):

      So double disadvantage for panda here.

      Additional thing to check related to ‘stream’ benchmark is -fprefetch-loop-arrays gcc option. It enables software prefetch instructions emitted by the compiler and greatly improves performance on such workloads. That is if the target processor does not have hardware prefetch. Cortex-A8 does not have hardware prefetch for sure (so no good performance for Efika without it). Cortex-A9 *may* have a working hardware prefetch unless it is misconfigured by the boot rom, bootloader or the kernel.

      Another thing regarding ‘steam’ benchmark is that its MB/s numbers are inflated compared to ‘memcpy’ benchmark (it counts both read and written data separately, so steam ‘Copy’ test shows 2x higher numbers than ‘memcpy’ for the same performance).

  6. Matija "hook" Šuklje Says:

    Thanks Raúl for your answers. 🙂

  7. Karel Gardas Says:

    Hi Raul, thanks a lot for the stream bench! Tegra2 looks really interesting. Karel

  8. Bel Says:

    I’ve tried Tomcat with some networked apps and i found it actually fine for its specs,
    open-jdk6, Tomcat5 on Sheeva ARM, 512M RAM,

    However, I’ve tried running the same set of apps on GuruPlug Server – PLUS, Sheeva’s cousin, and it was just crashing, “OpenJDK was slightly higher version”
    So I guess the open-jdk6 on ARM is buggy in certain functions regarding opening sockets, etc…I haven’t got to the bottom of it yet, but JVM just crashes!
    also, i didn’t find it so easy to downgrade JDK on this little beast

  9. BaboonS Says:

    Memory benchmarks for a OMAP 4460 – looks about 2x improvement over my rev ES2.1 pandaboard

    # ./memspeed_a9 8192k 8M
    offset 8388608, 0
    buffers 0x40001000 0x40801000
    copy libc 573 MB/s
    copy Android v5 575 MB/s
    copy Android NEON 604 MB/s
    copy INT32 523 MB/s
    copy ASM ARM 564 MB/s
    copy ASM ARM A9 586 MB/s
    copy ASM VLDM 64 599 MB/s
    read ASM ARM 1487 MB/s
    read ASM VLDM 1445 MB/s
    write libc 2480 MB/s
    write ASM ARM 2515 MB/s
    write ASM VSTM 2513 MB/s

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s

%d bloggers like this: