Mercury MultiCore Plusâ„¢ SDK

Discussion in 'CellPerformance@B3D' started by sevanig, Aug 4, 2007.

  1. sevanig

    Regular

    Joined:
    Jul 28, 2005
    Messages:
    254
    Likes Received:
    0
    Location:
    Sydney, Australia
  2. gokickrocks

    Regular

    Joined:
    Dec 19, 2002
    Messages:
    465
    Likes Received:
    1
    what in matlab is so computation heavy that you are doing now?

    if its computation heavy, you shouldnt be using matlab in the first place

    ...and what makes this turn the ps3 into a supercomputer?
     
    #2 gokickrocks, Aug 4, 2007
    Last edited by a moderator: Aug 4, 2007
  3. 22psi

    Veteran

    Joined:
    Sep 23, 2006
    Messages:
    1,365
    Likes Received:
    24
    Come on, stop the madness. :lol:
     
  4. randycat99

    Veteran

    Joined:
    Jul 24, 2002
    Messages:
    1,772
    Likes Received:
    12
    Location:
    turn around...
    I'm no math expert, but I have seen the matlab creations of others that seem to get quite hairy. Computational prowess could be quite a benefit when it comes to transformation of vast datasets, extensively iterative procedures, concurrent exploration of multiple scenarios, brute force approaches to automated parameter optimization on a particular mathematical model... These sorts of explorations are not necessarily requiring extreme precision, either (i.e, double precision performance). Single precision is quite adequate. It's just an extreme amount of math that needs be done to get from A to B.

    Typically, development of mathematical models take into account what is computationally feasible for the current, available hardware of the day. So that makes it possible, that formerly prohibitive techniques/procedures/scale of data now become plausible avenues when a new level of computational capability becomes available.
     
    #4 randycat99, Aug 4, 2007
    Last edited by a moderator: Aug 4, 2007
  5. gokickrocks

    Regular

    Joined:
    Dec 19, 2002
    Messages:
    465
    Likes Received:
    1
    dont know if you were for or against what i said from reading your post...

    but the reason why i said what i did about matlab is that its an interpretive language, so things that need iterating like loops just run slow...if you need something computationally intense, you are better off creating your own app using the c++ (or whatever language you use) math libraries that are available around the web
     
    #5 gokickrocks, Aug 4, 2007
    Last edited by a moderator: Aug 4, 2007
  6. patsu

    Legend

    Joined:
    Jun 25, 2005
    Messages:
    27,614
    Likes Received:
    60
    Matlab contains optimized, parallel matrix solver(s). It is fast. The interpretor is simply a layer on top. If your matrices are large enough, the overhead should be small. In fact, you can google for "Matlab vs C" to see users' comments. It is not as simple as you may think.

    As for what makes PS3 a supercomputer... I think people make that connection because it is indeed evolved from supercomputing concepts in the 90s. Large scale distributed applications like Folding@Home runs on it. In addition, there are supercomputers built using the Cell chip... although newer ones will be using later Cell revision. I also remember an article mentioning that Sony went through great effort to make sure PS3 can withstand PC-like usage condition (Always on, compute intensive, workhorse use cases).
     
    #6 patsu, Aug 4, 2007
    Last edited by a moderator: Aug 4, 2007
  7. gokickrocks

    Regular

    Joined:
    Dec 19, 2002
    Messages:
    465
    Likes Received:
    1
    the matlab vs c arguments usually boils down to dev time (and how good the dev is for that matter, because bad code is just bad code)

    i believe matlab is written in c, so its up to the dev whether or not the c code will run faster than the matlab code (algorithms and tuning to a specific architecture)

    as for "supercomputer" remark...all those arguments are fine but you are talking about clusters, whereas i am referring to one ps3, because using the cluster argument, any computer can be used for a supercomputer
     
    #7 gokickrocks, Aug 4, 2007
    Last edited by a moderator: Aug 4, 2007
  8. Carl B

    Carl B Friends call me xbd
    Moderator Legend

    Joined:
    Feb 20, 2005
    Messages:
    6,266
    Likes Received:
    63
    I don't think the goal here is to turn the PS3 into a supercomputer - rather, simply to release an SDK for it that allows for development and testing of code that will port seamlessly and/or scale up through multi-node machines.

    The story here is the SDK itself, rather than what the end result is on the PS3. The PS3 in this case is simply a cost-effective means to an eventual end.

    PS - Here's the link to the actual press release: http://www.mc.com/microsites/cell/PressReleases.aspx?id=10454
     
    #8 Carl B, Aug 4, 2007
    Last edited by a moderator: Aug 4, 2007
  9. patsu

    Legend

    Joined:
    Jun 25, 2005
    Messages:
    27,614
    Likes Received:
    60
    How does that contradict the press release ? The SDK is optimized for Cell (e.g., using intrinsics/SPU assembly), meaning the dev does not need to waste time building a good library. Also, performance is likely to be superior to stock workstation solution.

    Not really. If you're not careful, you may end up with a bad performing cluster or one that takes too much power as you scale up. There are good reasons for Cell to be in the world fastest supercomputer: http://news.bbc.co.uk/2/hi/technology/5322704.stm

    I know of a cluster that will cause a brown out in the neighbourhood when they try to max it out (It's not particularly fast compared to the bests).

    In addition, Cell's single node performance can be ridiculously good when the algorithm maps well: http://forum.beyond3d.com/showthread.php?t=39552

    Because the same CPU _and libraries_ run on PS3 and Cell workstations, you can write the app on a PS3 and time it in a cluster for performance. For an example, see standing ovation's post in another thread here. A solution typically starts its life on 1 node first anyway. The SDK is a cheap way for research projects to start and for the academia to spread parallel programming know-how.


    EDIT: standing ovation's post can be found here: http://forum.beyond3d.com/showpost.php?p=1044020&postcount=150
     
    #9 patsu, Aug 4, 2007
    Last edited by a moderator: Aug 4, 2007
  10. gokickrocks

    Regular

    Joined:
    Dec 19, 2002
    Messages:
    465
    Likes Received:
    1
    i think we are arguing 2 different points here, i wasnt trying to contradict the press release, i was rationalizing with the original poster...my point was why use matlab if you have the math libraries from the sdk for something thats computationally intense, matlab will still use its own libraries whereas the libraries from the sdk will be optimized for cell so for serious number crunching you should make your own app that utilizes those optimized libraries

    i believe that cell can be a great basis for a supercomputer, but thats not what im arguing

    all supercomputers start life on 1 node

    ...
    again, im not arguing that ps3/cell can be a great BASIS for a supercomputer, im saying that the sdk will not automagically turn a ps3 into one as the original poster implied...just installing it wont allow matlab to run at supercomputer speeds

    i am in now way implying that cell is not good, in fact i would love to play around with the sdk's math libraries
     
    #10 gokickrocks, Aug 4, 2007
    Last edited by a moderator: Aug 4, 2007
  11. patsu

    Legend

    Joined:
    Jun 25, 2005
    Messages:
    27,614
    Likes Received:
    60
    Why can't someone optimize/extend MATLAB instead of doing a specialized app ? Productivity gain is useful in the realm of supercomputing too. The poster is just speculating that one can speed up MATLAB using the SDK.

    In your post, "what in matlab is so computation heavy that you are doing now?"... if the matrices involved are huge, and the problem is NP complete, then MATLAB will be compute intensive.

    Even on Blue Gene, installing _any_ SDK won't automagically turn out good performance. The point is the Mercury SDK enables the developers to compute difficult problems at speed much faster than traditional CPUs, and at relatively low power consumption due to Cell's design. That is all there is to it.

    Sure... I look forward to your results.
     
  12. gokickrocks

    Regular

    Joined:
    Dec 19, 2002
    Messages:
    465
    Likes Received:
    1
    well i should have said app/dll since most people use MEX in matlab, so ill give you that

    i agree, hence why i mentioned dev time in an earlier post

    i may have assumed more from the original post due to the reference to a supercomputer

    huge matrices may not be as viable on a ps3 compared to a pc,...a pc has a better chance at working on a huge matrix than a ps3, it may take longer but it will get computed, it has nothing to do with cell not being capable of it, its that the ps3 is limited by memory...but then again those would be very big matrices

    again, not disagreeing with this
     
    #12 gokickrocks, Aug 5, 2007
    Last edited by a moderator: Aug 6, 2007
  13. patsu

    Legend

    Joined:
    Jun 25, 2005
    Messages:
    27,614
    Likes Received:
    60
    If we keep making generalized assumptions like this, Folding@Home may not run on PS3 today.

    For some of these problems, PS3 should be a good starting point. Once the solution is proven, the "same code" can be run against a bigger problem on a larger machine or a cluster (of PS3s).
     
    #13 patsu, Aug 6, 2007
    Last edited by a moderator: Aug 6, 2007
  14. Todd33

    Veteran Banned

    Joined:
    Jan 22, 2007
    Messages:
    1,066
    Likes Received:
    7
    Location:
    CA
    http://www.mc.com/ps3/default.aspx

     
Loading...

Share This Page

  • About Us

    Beyond3D has been around for over a decade and prides itself on being the best place on the web for in-depth, technically-driven discussion and analysis of 3D graphics hardware. If you love pixels and transistors, you've come to the right place!

    Beyond3D is proudly published by GPU Tools Ltd.
Loading...