Supercomputer-on-a-chip microprocessor revealed

Discussion in 'Overclocking & Hardware' started by Warpy, Feb 8, 2005.

  1. Warpy

    Warpy Member

    Joined:
    Nov 25, 2004
    Messages:
    3,948
    Details of a novel microprocessor design that could supercharge many computing applications were released at an industry conference in California, US, on Monday.

    The microprocessor architecture - known as Cell - will appear in the Playstation3 games console, scheduled for release in 2006. But experts say it could ultimately find its way into many home entertainment devices, high-end computers and even supercomputers.

    Details of the chip were disclosed at the International Solid State Circuits Conference in San Francisco, stirring debate over the possible implications for the computer industry.

    Developed jointly by IBM, Sony and Toshiba, the microprocessor is fundamentally different from the chips that power most computers today. It incorporates eight separate processing cores, or "synergistic processing elements", which are capable of communicating with one another at high-speed. A standard chip has single, larger processor.

    The eight cores give the chip the ability to run different software programs simultaneously and to divide up processing tasks more efficiently. It will also be able to run several operating systems, such as Linux or Windows, in tandem.

    Source
     
    Last edited: Feb 8, 2005
  2. shift

    shift Member

    Joined:
    Jul 28, 2001
    Messages:
    2,937
    Location:
    Hillcrest, Logan
    As well as the eight vector cores, theres a full 64bit Power-derived core, with SMT and VMX.
    The vector cores all look to be VMX like, enhanced with Double Precision float support.

    EDIT: Apparently I was wrong about the vector cores, no DP support, and missing some of the capabilities of VMX. Seems to have the important permute unit though.
     
    Last edited: Feb 8, 2005
  3. csimpson

    csimpson Member

    Joined:
    Jan 24, 2005
    Messages:
    7,644
    Location:
    Sydney
    or as we like to say in the industry,

    zing.

    sounds cool.


    C
     
  4. Myne_h

    Myne_h Member

    Joined:
    Feb 27, 2002
    Messages:
    8,200
    intel already has this dual os crap in the works. multicores are soon to appear on shelves from both companies and they're both moving 64bit. There are also several examples of supercomputers that use x86 chips - does that mean I have part of a supercomputer?

    Gimme a H
    Gimme a Y
    Gimme a P
    Gimme a E

    Actually, the biggest change with this arcitecture is probably in the organising of the processing and the compiling of programs. Its existance might force a re-examination of several inefficiencies of the x86 architecture and hopefully lead to more improvements.
    It'll have an impact for sure, but I think its distinctiveness will just be assimilated :)
     
  5. shift

    shift Member

    Joined:
    Jul 28, 2001
    Messages:
    2,937
    Location:
    Hillcrest, Logan
    Yes, a very, very small part.
    Intel has no announced plans for anything like this (Peak 250+ GFlops per chip - 60 of these could theoretically get you into the top 10, compared to the 2500 P4s required).
     
  6. GreenBeret

    GreenBeret Member

    Joined:
    Dec 31, 2001
    Messages:
    19,370
    Location:
    Melbourne
    If they indeed can produce something that powerful, I wonder how much power it will need to run and how the hell they can fit it in a console.
     
  7. shift

    shift Member

    Joined:
    Jul 28, 2001
    Messages:
    2,937
    Location:
    Hillcrest, Logan
    It's around 230 million transistors, around what current GPUs are at. That's a fair bit for a console to handle, especially when the PS3 is supposed to have multiple cells plus a GPU.
     
    Last edited: Feb 8, 2005
  8. Deeman

    Deeman Vagrant.

    Joined:
    Jun 28, 2001
    Messages:
    8,727
    Location:
    Melbourne
    Since when has gigaflops been a useful measure of CPU performance?

    Cell is cool, make no mistake.

    However, cell isn't the leap that people are hoping for, and it's not going to be easy to program for (neither was the original Playstation & PS2, so no change there).

    A decent primer on cell can be found on Ars... (Hannibal will be expanding on it in the near future, so stay tuned)

    http://arstechnica.com/articles/paedia/cpu/cell-1.ars

    Warpy, care to give us your opinions? You're posting verbatim a lot of press-releases without any further discussion - Agg posts the press releases regularly in a single update...
     
    Last edited: Feb 8, 2005
  9. bsbozzy

    bsbozzy Member

    Joined:
    Nov 11, 2003
    Messages:
    3,924
    Location:
    Sydney
    That press release said it could run windows, is this true, casue if it could run windows, it shouldnt be that hard, but i am seriously unsure on this?
     
    Last edited: Feb 8, 2005
  10. shift

    shift Member

    Joined:
    Jul 28, 2001
    Messages:
    2,937
    Location:
    Hillcrest, Logan
    The official press releases don't mention Windows at all.
    It wouldn't be impossible for MS to port windows to it, but they have no reason to at this point.
     
  11. noboundaries-au

    noboundaries-au Member

    Joined:
    Jan 9, 2005
    Messages:
    7,571
    Location:
    Canberra, ACT

    Wouldnt it have to run x86 architecture so we can use existing programs. without an emulator that is (thinks of XP dos support :upset: )
     
  12. Flamous

    Flamous Member

    Joined:
    Dec 6, 2003
    Messages:
    779
    Location:
    Brisbane
    Not only are they multi cored but multi threaded as well. Hyperthreading is two threads running on the processor whereas these will have 4 threads running on each core. With good programming these should be more efficient, especially as they will most likely share cache.

    I also read somewhere that the PS3 will have 2 Cell based processors which when combined to work together produce the equivalent processing ability of a 250Ghz Pentium 4. Sounds a little outrageous to me but i guess we will see.
     
  13. Hairy_Harold

    Hairy_Harold Member

    Joined:
    Sep 11, 2004
    Messages:
    1,301
    Location:
    Ipswich
    well god look at the size of the xbox :| you call that a console still, looks like a 1980's vcr to me with some modern style.
     
    Last edited: Feb 8, 2005
  14. Deeman

    Deeman Vagrant.

    Joined:
    Jun 28, 2001
    Messages:
    8,727
    Location:
    Melbourne
    Pretty much. There was a version of NT4 available for Alpha, and various more recent windows versions for IA-64 (Itanic), but application support is very scarce.

    As for the 'processing power of a 250GHz P4' - it's an impressive technical feat, but numbers like that are generally false, or pie in the sky estimates, taking the perfect scenario where the P4 struggles, and the massive symmetric processing of something like the cell benefits greatly.

    As people have said, massive multithreading is great, if your application supports it, and you code it that way.

    Massive SMT-capable systems are useless if your codepath requires the results of previous calculations, or is inherently a single-threaded application.

    Dual core is the same - until developers start to think in a true multithreaded manner, all the transistors in the world aren't going to help you. Great technology, with great potential, but it really requires a totally different way of thinking to the way that programs and programmers have worked for the entirety of the modern computer industry.

    Video cards already do massive parallelism (16 pipelines), but it's easy when you know the exact order in which you perform your operations, which you do with something like rendering graphics. Rendering a scene is always the same process - the shaders used may be different, but everything before and after is identical, and you don't care what the other pipes are doing (unless you go multipass, which hurts performance a lot).

    It's the way of the future, but it's not going to happen straight away. Watch as most current applications and tasks don't see a whole lot of benefit from dual core. Unless the application is much faster under SMP systems already, dual cores will do 4/5ths of stuff all for it - and that includes pretty much *all* modern games.

    Edit: When people start throwing 'gigaflops' and the like at you as a performance measure, you know you're being thrown marketing. Gigaflops is a totally synthetic measure of performance, and really tells you nothing about relative performance. Just talk to the other great Gigaflop marketer (and the 'zOMG it's a soopercomputer on yuor DESK!' people) out there at the moment - Apple.

    IBM make both the G5 and Cell, so don't be too suprised they're spinning the same line. I'll bet there's more than a hint of Altivec in cell...
     
    Last edited: Feb 8, 2005
  15. shift

    shift Member

    Joined:
    Jul 28, 2001
    Messages:
    2,937
    Location:
    Hillcrest, Logan
    There was also an NT4 version for PowerPC.
    A lot of benchmarks are totally synthetic. Gigaflops, specifically actual achieved linpack rate, is used often to rank supercomputer performance.
    The artifical 'peak gigaflops' figure used here is about the only performance calculation that can be made for a raw cpu without the rest of the system being a factor - so to compare two CPUs of differing architectures, rather than two systems, it's all we really have.
    And it will make the internetz fasterz.
    No wait, that's Intel. Marketing is marketing, and should be ignored by anybody with half a clue.
    Yes, the Cell control processor contains a fully functional Altivec, aka VMX, unit as has already been stated.
     
  16. munka

    munka Member

    Joined:
    Aug 1, 2001
    Messages:
    2,460
    Location:
    Perth
    They havent described exactly how it works, but lets look at it this way.

    If you have several small cores you can do cool things like speculative branch prediction. Where instead of guessing what the branch should be you just run two threads then ditch the one that wasnt right.

    Strongly threaded apps in a hyperthreaded situation leave less problems with cache misses.

    And in general if instead of running four instructions through at once with a hugely complicated scheduling unit you can just run one or two, and have more cores.

    My computer structures lecturer claims the current PC core(cache excluded) is half scheduler, and the other stuff is another 15%, leaving only 35% for register files and execution units.

    So lets say each cell uses 1/4 the transistors it makes sense that they can fit 9 on a chip instead of intels 2.

    With cache sharing this should be very efficient. Each core might only be as good as a Via C3 or something but if the PS3 has 18 cores its gonna be a pretty decent chip.
     
  17. Varuna

    Varuna Member

    Joined:
    Sep 20, 2004
    Messages:
    1,046
    Location:
    Sydneys Northern Beaches.
    Power of a 250Ghz Pentium 4 for highly vectorisable tasks rarely performed on a PC, file, web or database server.

    Someone will probably think of a way to offload discrete fourier transforms and the like to a cell card with a few hundred MB of fast memory jammed on the PCI-e bus. Might give those who rushed into SLI something to use the extra 16x slot for. :rolleyes:
     
    Last edited: Feb 9, 2005
  18. Devilsmurf

    Devilsmurf (Banned or Deleted)

    Joined:
    Jul 9, 2003
    Messages:
    1,422
    http://arstechnica.com/articles/paedia/cpu/cell-1.ars/2


    The SPE's ISA, which is not VMX/Altivec-derivative (more on this below)...Finally, the instruction set for the SPEs is not VMX compatible or derivative, because its execution hardware doesn't support the range of instructions and instruction types that VMX/Altivec does.
    (Quoted from Ars Technica)
     
  19. shift

    shift Member

    Joined:
    Jul 28, 2001
    Messages:
    2,937
    Location:
    Hillcrest, Logan
    Yes, the SPEs do not use Altivec. The PPE however does.
     

Share This Page