Saturday 3 August 2013

The big.LITTLE ODROID-XU

One of the community members over at POGS just mentioned that the new ODROID-XU is now available form Hardkernel. Check it out here: http://hardkernel.com/renewal_2011/products/prdt_info.php?g_code=G137510300620

Quite the beefy little ARM unit hosting the new Exynos 5 with 8 cores (4 usable at a time). Essentially it switches between 4 cores that are for light tasks and 4 which are for heavy tasks. Not sure if it's possible to use them all at once, haven't looked into this yet. The whole switching thing is ARM's "big.LITTLE" trademarked design in an attempt to be more energy efficient on mobile devices. The GPU is compliant with OpenCL 1.1 embedded platform profile.

I'm having a serious battle with myself over if I should get it or not. I currently have the ODROID-U2, however, I would like to get my hands on this one because of the GPUs OpenCL 1.1 compliance, albeit only embedded profile as opposed to full. It would allow me to test some OpenCL code I've been playing with for BOINC on an ARM based device.

The problem is that I'm in no position to throw money at new toys. Oh, what to do? Maybe I should wait for something in the same class packing a GPU with full profile OpenCL 1.1 support.

16 comments:

  1. I'm waiting for the Parallella to arrive. They have OpenCL for their Epiphany chip which might help your craving for OpenCL. Let me know if you're interested and I might see what I can do :-)

    ReplyDelete
    Replies
    1. I'm also waiting for my Parallella to arrive and I may grow old waiting for them to come :)! The thing is the Epiphany's don't do full profile OpenCL and only do single precision floats. I will still be testing on them when they do come. I think I ordered two of them...

      I'm waiting for a dev board style device with Mali GPUs that supports full profile OpenCL (including double precision floats). This will be sweeet.

      Delete
    2. Does the http://www.qualcomm.com/snapdragon/processors/800 fit the bill IRO full profile OpenCL via the Adreno 330 GPU?
      Not a dev board, I know, but a teardown would do that.

      Like the Tegra 4 Nvidia Shield board revealed here http://www.ifixit.com/Teardown/Nvidia+Shield+Teardown/16212/3
      :-)

      Delete
  2. [quote]a dev board style device with Mali GPUs that supports full profile OpenCL (including double precision floats). This will be sweet[/quote]

    Me too. Please let us know if/when one appears. ARM boards are OK but OpenCL -erm - opens a whole new level of processing power.

    I cancelled my order for 2 XUs due to the CCI 400 bug.
    The XU has the CCI-crippled 5410 SoC that sold poorly in phones with that bug whereas Samsung (AKA Hardkernel?) quickly released the bugless 5420 SoC so I'll wait for that to appear.

    As for the Parallela ... maybe kickstarter customers will get them "soon" but I just read the following:-
    http://forums.parallella.org/viewtopic.php?f=10&t=528#p3211
    I'm not making any plans.

    ReplyDelete
  3. More Parallela stuff ... but as a Business Analyst by trade, funding clearly wasn't thought through properly.

    From their website on August 21, 2013:-

    "the Gen0 Parallella board is definitely useable, but it is clearly not the product we set out to build when we started the Parallella project.

    The Gen1.0 board is back from the fab and the board assembly starts tomorrow ... we will be able to show conclusively that we have solved the overheating issue that plagued the Gen0 design."

    Two things:-

    The original Kickstarter project ended far too soon. Given just a few more weeks the $3m/64-core device targets were achievable IMHO.
    Commendable transparency in all aspects of the Parallela/Adapteva build. Others would do well to follow this example e.g., RPi board dimensions/absence of support for Android, ODROID-XU crippled Exynox 5410 SoC via CCI-400 bug and lots more stuff that is waiting for the future to judge.

    ReplyDelete
    Replies
    1. Saw this and thought - Daniel might be interested.

      http://www.parallella.org/2013/09/10/explorations-in-erlang-with-the-parallella-playing-with-pixels/

      FWIW my eyes glazed over at "Erlang" but I kept on trying ... ;-|

      Delete
    2. Hmmm, very interesting. Sorry I've been a bit lazy lately with blogging!

      Delete
  4. Hey Daniel.

    FINALLY a dev on the XU has admitted the big.LITTLE SNAFU (link available, ask if you're interested).
    Parallela - words fail me. Hints of massive problem/s and details TBA "soon", pics of '000 of cores, prices:- cluster $179 greater than the cost of 4 boards, Epiphany-III 16-core chips @$75 ea ... aaargh. What is going on there?

    ReplyDelete
  5. Does this stuff mean anything to you ... https://lkml.org/lkml/2013/10/1/264 ???
    If yes, help me decide whether I should buy a couple of XU's.
    TIA, Ray

    ReplyDelete
  6. Hmmm, I'm not too sure to be honest. I haven't got the device to test that code but I don't like the idea of having the slow and faster cores running at the same time. I feel like it would just be more problems than it's worth, especially for crunching.

    ReplyDelete
    Replies
    1. Just posted this elsewhere ... what do you think?

      I did reconsider purchasing an XU or 2 a couple of days ago following that link which I thought was a final version.
      Then I looked again, tonight.

      First off, I'm not a coder (any more) and have never done C therefore also no LK development.
      But what little I understand of that link & subsequent comments by Linux kernel development contributors it appears as though the big.LITTLE and (probably more importantly?) the CCI implementation has been a properly bad idea.

      If it were possible to wind back the clock then I'd say everyone would be better off with the well-tried multi-core hardware/software that has been around for years, especially since, as has been said before, any task lights up the A15 cores anyway.

      Taking a step back from where we are now ... given the time LK development will take to sort out the big.LITTLE/CCI situation then such LK work is likely to be left far behind by better, cheaper multi core hardware anyway. And very, very quickly.

      Put another way, if "I were you, I wouldn't start from here".

      The evidence is already out there today, with quad core A15 tablets, true-octa (A9) tablets all sporting a working Android OS available right now and 12/16 core A15 devices available when the hardware is finalised.

      To summarise:- in the UK we have an expression that I have been holding back saying IRO big.LITTLE for some time ... this is like "flogging a dead horse". Eventually, the XU might get it's 8 cores up and running but by then things will have moved on. Quite far.

      Delete
  7. I'm intrigued ... why would there be "more problems"?
    As I see it from my simple user perspective ... tasks being processed, specifically crunching, would simply take longer with a slow core and complete more quickly with a faster core ... the tasks are independent and don't interact e.g., 1xWUprop & 2xEnigma & 1xMW GPU at the same time on my ancient P4/HD3650 AGP/XP running in only +/-3gb memory work just fine?

    ReplyDelete
  8. PS:- like me I guess you're waiting for the TSMC 16nM FinFET ARM A57 that "promises three times the CPU power of current chips like Samsung's Exynos 5" (A15/XU)? Let me know if you spot one - cheap :-)

    ReplyDelete
  9. pps:- 1st to 1 million credits on an ARM device - see http://nativeboinc.org/site/detail_host/157557 (ODROID-X)

    TBH it's the 3rd ... as I have 2 other devices well past 1 million credits on the same ARM device, just that NativeBOINC/BoincSTATS can't "see" the same device if it has two frequencies (1704MHz & 2000MHz).

    ReplyDelete
  10. amendment to above (link only) ... see http://nativeboinc.org/site/host_stats#order=credits

    (for more details etc see http://forum.odroid.com/viewtopic.php?f=29&t=2468)

    ReplyDelete
  11. Users report that the ODROID-XU seems to run hot???
    See Hardkernel thread http://forum.odroid.com/viewtopic.php?f=65&t=1905&start=20 e.g., lizardmech says "it will reach at least 60c under load regardless of how good air cooling is"

    Got yours yet?

    ReplyDelete