Apple's High-Water Mark?
Subject:   Cell Processors
Date:   2006-03-23 18:02:11
From:   malibu
While the cell processor marketing is good the reality is still in question. Apple was offered the Cell and turned it down after evaluating it.

The big issue, one very similar to the Itanium processor are the compilers and tools. In order to maximize the performance on the Cell archtecture you have to program in a different model then most programs are written. More concerency and specialty options must be used to maximize throughput. IBM has admitted as much as they are working on a project called "Octopile"that will take programs written in standard programming languages and standard models to get them to perform optimally on the Cell architecture. I am sure this will take some time to mature.

These are the same problems that exist with the Itanium. It currently does not support an on board instruction re-ordering unit and thus all branch predictions and optimization MUST be done at compile time to avoid pipeline issues.

The G5 suffers from this in some ways as well. The fact that the intel architecture has had over 10 years of compiler development in terms of gcc alone and that optimizations and intrinsic library support is significantly more mature provides for huge gains in performance. The G5 also lacks a good set of support chips surrounding it. Much of the support ASICs in Apple's boxes were developed by Apple. With the intel architecture much of the supporting architecture is done and Apple can concentrate on other areas.

Full Threads Oldest First

Showing messages 1 through 3 of 3.

  • Cell Processors
    2006-03-24 17:58:21  AdrienLamothe [View]

    Initial reports of Cell demos have been very favorable, but only time will tell. Early demos do tend to be tightly controlled affairs.

    I don't expect the different programming model to be a show-stopper.

    Excellent point about the compiler optimizations and chipsets. A company with the scope, size and organizational capacity of IBM should be able to overcome those types of problems.

    • Cell Processors
      2006-03-25 05:50:24  malibu [View]

      A show stopper no but a significant obsticle yes.

      A brief description of the coding issues can be found here

      I would not be so sure about IBM overcoming this issue quickly and confidently. They had same said resources and look where the G5 ended.
      • Cell Processors
        2006-03-25 09:39:49  pquam [View]

        I agree that this won't be solved overnight to
        anyone's long-term satisfaction. Concurrency
        is a hot topic in CS, and has been since the 80's
        or before. So, in a sense, it has been the
        subject of a generation's doctoral theses, and still, I would argue, it's not mature, and few
        people would disagree. There have also been
        some commerical failures, as in the 5th generation
        computer project--which was not a failure in
        terms of teaching us about what not to do.

        I don't think this negates the usefulness of the cell processor though. In the worst case, ignoring the SPEs, you have a simple risc-based dual-threaded processor clocked up to about 4ghz, that's compatible with all the powerpc code. That means that there is a huge code-base that will work already with it. This may or may not be enough to make this a competative product.

        There will, of-course, be sub-optimal usage of
        the SPE's for the forseeable future with most
        applications. This is not necessarily a bad
        thing. What dedicated graphics has taught us
        is that you can take certain processor intensive
        tasks, and off-load them relatively cheaply
        to dedicated chips. Even selective usage
        of the SPE's could make a huge performance
        difference on certain processor-intensive tasks.

        What I see as the key feature of the cell
        processor for 2006 is not its parallel
        execution, but its I/O bandwidth. It's an
        incremental, but very significant advance on the
        real bottle-neck in modern processors: bandwidth.

        Bandwidth should be reason enough to make Cell
        successful until the tools for painless and
        automatic compiling can be developed.