• Yokemate of Keyboards
    Yokemate of Keyboards
    Andreas_Wolf
    Posts: 12079 from 2003/5/22
    From: Germany
    > multithreaded

    Nice to see that Freescale realized their SMT plans.

    > http://media.freescale.com/phoenix.zhtml?c=196520&p=irol-newsArticle&ID=1576370
    > Seems they really value Altivec again.

    Yes, it seems that QorIQ AMP is the official moniker of what has so far been on Freescale's revised roadmap as "QorIQ T series" sporting "improvements incorporated in the updated version" of AltiVec.

    > Question is why they ditched it first then?

    It's surely been a costly and time consuming undertaking to integrate AltiVec with a core derived from e500, so Freescale had hoped to get along without that after they decided to not develop e600 any further. What they obviously didn't anticipate though is that many of their customers very much depend on the presence of a powerful SIMD unit and thus were about to abandon Freescale as a supplier and go to Intel instead. I'd regard the inclusion of AltiVec into QorIQ as a measure to some kind of damage control. Needless to say that I appreciate this wise decision to re-introduce AltiVec of course.

    Edit: statement from Glenn Beck, marketing manager for aerospace, defense and single board computing within Freescale's Networking Processor Division, regarding the ditching of AltiVec (from the AltiVec video linked below at 3:16):
    "When it was not on the roadmap we did it for a reason. It was because of market pressures. But those pressures have changed. We're clearly seeing the need in terms of printing and imaging, video surveillance, analytics inside of surveillance, and even in the networking world a need for a SIMD engine that complements the scalar processor that is already out there in QorIQ."

    I wouldn't know what those "market pressures" that have changed were, though.


    Some more resources on the e6500 core and the QorIQ AMP series:

    http://www.freescale.com/webapp/sps/site/overview.jsp?code=QORIQ_AMP (product page)
    http://www.freescale.com/files/32bit/doc/fact_sheet/QORIQAMPFS.pdf (QorIQ AMP fact sheet)
    http://www.freescale.com/files/netcomm/doc/fact_sheet/QORIQOV.pdf (QorIQ P and AMP fact sheet)
    http://www.youtube.com/watch?v=A3THXh9qD3o (QorIQ AMP introduction*)
    http://www.youtube.com/watch?v=YGM7femV4uA (QorIQ AMP overview**)
    http://www.youtube.com/watch?v=wUW74jSIYXg (QorIQ AMP overview)
    http://www.youtube.com/watch?v=lOJP2YRr6Uc (e6500 core overview***)
    http://www.youtube.com/watch?v=6_cbntPxbck (AltiVec capability)
    http://www.youtube.com/watch?v=Vllchtb5mos (e6500 core SMT capability)
    http://videos.cache.magnify.net/60X0NG37F35BHBT2-FreescaleEmbargo0621_483_336_384x216.mp4 ****

    * Funny claim at 0:42: "Up to 100 GHz Performance" ;-)
    ** At 1:05: "...the e6500 [...] increases the core performance alone by 2x over the previous generation. [...] We've also increased the frequencies to up to 2.5 GHz..." Is the e5500 repudiated here?
    *** That one made me almost fall off my chair. Starting at 3:35 the engineer mentions that dual-threading increases core performance by about 60% while increasing power consumption by only about 30%, from which the interviewer (Freescale's John Dixon, Power and DSP products marketing manager) calculates a 50% relative power saving. Please someone say he didn't really divide 30 by 60 to come to his conclusion. Given my math isn't completely off I'd calculate 1-(1.3/1.6) and end up with a 20% relative power saving. Or did I miss something here?
    **** This calculating expert John Dixon from above claims at 0:57: "We also have the reintroduction of AltiVec, which is an enhanced version of our AltiVec that we had in our PowerQUICC products in our previous generations." WTF?


    From CEO Rich Beyer's keynote:

    "QorIQ AMP [...] is expected to deliver 4x the performance of our current family of products at one-half the power consumption." (page2)

    This is of course bollocks. What he probably meant to say is that it delivers 4x the performance at one-half the power consumption *per performance*, which amounts to 2x the power consumption at 2x the power efficiency. 4x the performance at one-half the power consumption would actually amount to *8x* the power efficiency.
    This specifically means that the T4240 consumes 60 watts (and not 15 watts as indicated by Beyer) to deliver 4x the performance of the P4080 at 30 watts.

    "we are adding virtualization capability" (page 3)

    He seems to be unaware that this has already been in there since the e500mc core.

    [ Edited by Andreas_Wolf 25.06.2011 - 15:41 ]
  • »21.06.11 - 12:41
    Profile