Multiprocessors...

Posted:
in Future Apple Hardware edited January 2014
Well, I've seen more and more threads about MP G4 (quad, octo-...), but I think it isn't realistic.



Why ?

IIRC, in order to have efficient MP configurations (# proc &gt; 2), processors must implement MERSI (modified, exclusive, reserved, shared, invalid) protocol to reduce the overhead while transferring data between the processors. ( few details <a href="http://maccentral.macworld.com/storyforum/forums/2001/07/30/hypertransport/?read=35"; target="_blank">here</a>.)



7400 and 7410 proc. implements full 5-state MERSI protocol, but according to <a href="http://e-www.motorola.com/brdata/PDFDB/docs/MPC7450TS.pdf"; target="_blank">Moto</a> (p.8), 745X processors only implement MESI protocol ( lacking the most important "Reserved" state ).



According to Bad Andy's posts [PowerPC/Altivec guru] at <a href="http://arstechnica.infopop.net/OpenTopic/page?a=frm&s=50009562&f=8300945231"; target="_blank">arstechnica</a> (sorry, can't remember the exact link), making a quad-G4 using MESI protocol wouldn't be as useful as we hope, because of the additional overhead, and because designing such a rig would be a nightmare.



More over, current implementations of the G4 by Apple are actually memory bound when using Altivec. Even with DDR, dual G4 are limited by the poor memory subsystem.



So keep expectations low !



**CONFIRMED** <img src="graemlins/smokin.gif" border="0" alt="[Chilling]" />

Nothing more than a dual G4 ( perhaps at 1.4 or even 1.5GHz ) but certainly not a quad or an octo-processor Mac.
«1

Comments

  • Reply 1 of 39
    leonisleonis Posts: 3,427member
    I thought 7455 already has solved that "R" problem



    Anyway...the next batch of G4s should be 7470 and 7460
  • Reply 2 of 39
    hmurchisonhmurchison Posts: 12,419member
    I doubt that Apples yields would allow for Quads and certainly not Octo Processor systems. I'd be happy with Dual and a fast subsystem
  • Reply 3 of 39
    franckfranck Posts: 135member
    [quote]Originally posted by Leonis:

    <strong>I thought 7455 already has solved that "R" problem



    Anyway...the next batch of G4s should be 7470 and 7460</strong><hr></blockquote>



    Of course, but Moto downgraded the G4 from MERSI -&gt; MESI, when comparing 74xx to 745x.



    Do you really think Moto will reengineer G4-next-rev. to allow MERSI protocol ?

    Think "embedded market"
  • Reply 4 of 39
    programmerprogrammer Posts: 3,457member
    Lay the blame where it belongs, please: Apple's memory subsystem is actually quite good. The problem is that the current G4 is limited by its 64-bit wide 133 MHz MPX bus to a maximum throughput of ~850 MB/sec. This is a function of the processor not the memory subsystem.



    Otherwise I agree with your post. This could change very quickly if the next PowerPC chip (whether it is a modified G4 family member, or a "G5") includes a faster bus (or on-chip memory controller) and full MERSI support.
  • Reply 5 of 39
    eskimoeskimo Posts: 474member
    Who needs MERSI when you can have MOESI
  • Reply 6 of 39
    programmerprogrammer Posts: 3,457member
    [quote]Originally posted by Eskimo:

    <strong>Who needs MERSI when you can have MOESI </strong><hr></blockquote>



    Okay, its late and I'm tired. I don't get it, you'll need to explain the joke.
  • Reply 7 of 39
    franckfranck Posts: 135member
    Well, I'm in the same boat
  • Reply 8 of 39
    hmurchisonhmurchison Posts: 12,419member
    MOESI cache

    "Modified Owner Exclusive Shared Invalid" or MOESI, Cache Coherency Protocol, keeps track of data in CPU caches; identifies when data from one CPU is needed by another, and when data is shared between CPUs; and effectively reduces memory traffic, increasing available bandwidth.
  • Reply 9 of 39
    programmerprogrammer Posts: 3,457member
    [quote]Originally posted by hmurchison:

    <strong>MOESI cache

    "Modified Owner Exclusive Shared Invalid" or MOESI, Cache Coherency Protocol, keeps track of data in CPU caches; identifies when data from one CPU is needed by another, and when data is shared between CPUs; and effectively reduces memory traffic, increasing available bandwidth.</strong><hr></blockquote>



    So its pretty much the same thing as MERSI, just using AMD terminology...?
  • Reply 10 of 39
    franckfranck Posts: 135member
    What has MOESI to do with AMD ?
  • Reply 11 of 39
    A good question to ask is 10.2 built to work with more than 2 processors? In 10.1.5 and lower the software was set to two or less processors. You could read in a part of the Unix side as to the build type and how many CPU's it would work with.

    Is this road block still in place with the current beta builds of 10.2?

    If it is than only dual procesors are in the cards, If the limit is not there who knows what apple will place in the box.



    A heat sink talked about in the prototype box seems to be to big for just a single or dual processor system.



    Later Steve
  • Reply 12 of 39
    see my post in the New PowerMac pictures post...



    "Re: multiple processors and big heat sink.



    As much as the notion of a quadG4 sounds like the DB, the current Kernel of Darwin only supports 2 processors.



    On the other hand, the new PowerMacs are rumored to be released with 10.2 which maybe has a revised kernel. Hmmm. I think we would have heard about that by now. Enough people have played with Jaguar preview to have noticed and it is unlikely to be the kind of revision that is added at the last second before the product is shipped.



    (Er, off the top of my head, I typing hostinfo at command line will give you information like Kernel version (current release is 5.5) and number of processors supported.)"



    So? Someone with a beta of Jaguar want to be so kind as to give it a try?



    CGI
  • Reply 13 of 39
    jcgjcg Posts: 777member
    [quote]Originally posted by Franck:

    <strong>



    Of course, but Moto downgraded the G4 from MERSI -&gt; MESI, when comparing 74xx to 745x.



    Do you really think Moto will reengineer G4-next-rev. to allow MERSI protocol ?

    Think "embedded market"</strong><hr></blockquote>



    If Apple specified that this is what they wanted, then Moto would probably put it back in, after all they do have that part of the chip designed, and I would imagine that it would be relatively simple to put it back in.
  • Reply 14 of 39
    jerombajeromba Posts: 357member
    Mach kernel version: Darwin Kernel Version 6.0:

    Sat Jul 13 00:40:35 PDT 2002; root:xnu/xnu-333.obj~1/RELEASE_PPC





    Kernel configured for up to 2 processors.

    2 processors are physically available.

    Processor type: ppc7450 (PowerPC 7450)

    Processors active: 0 1

    Primary memory available: 1536.00 megabytes.

    Default processor set: 40 tasks, 91 threads, 2 processors

    Load average: 0.00, Mach factor: 1.99
  • Reply 15 of 39
    [quote]Originally posted by jeromba:

    <strong>Mach kernel version: Darwin Kernel Version 6.0:

    Sat Jul 13 00:40:35 PDT 2002; root:xnu/xnu-333.obj~1/RELEASE_PPC





    Kernel configured for up to 2 processors.

    2 processors are physically available.

    Processor type: ppc7450 (PowerPC 7450)

    Processors active: 0 1

    Primary memory available: 1536.00 megabytes.

    Default processor set: 40 tasks, 91 threads, 2 processors

    Load average: 0.00, Mach factor: 1.99</strong><hr></blockquote>



    Ok so the current Beta is only set up for two processors. Any idea how hard it would be for apple to change that before the final build of 10.2 goes to press?
  • Reply 16 of 39
    franckfranck Posts: 135member
    I'm afraid this won't change anytime soon.
  • Reply 17 of 39
    yevgenyyevgeny Posts: 1,148member
    As much as we would all like to see quad G4's, I think that we all understand that such a system would be crippled by its bus and that it could not get the performance that users would want. Unfortunately, the things that you would want quad G4's for are the things that eat up the bandwidth.



    Dual G4's are far more reasonable and I think that a dual 1.4 G4 with enough bandwith to keep the altivec units fed would beat any other machine out there in the standard video/photo benchmarks. If Apple came out with Dual 1.4's and a bus that worked with them, then they would be back in the speed race.



    I can see Apple making the entire Pro line a lineup with multiprocessor boxes. This is probably the fastest way to make up for the performance gap.
  • Reply 18 of 39
    aphelionaphelion Posts: 736member
    MERSI ME!



    I'd rather have 8 (or even 4) 7410's @ 500 Mhz than 2 74xx @ 1.5 Ghz. My Sig says it all, and I have had it on this forum since Jan 2000. Apple has always had the capability to win the Mhz Wars outright. They have chosen not to all this time and unfortunately will continue to lag behind by choice, not by any technical limitations set by Motorola or anybody else.
  • Reply 19 of 39
    ssmurphyssmurphy Posts: 40member
    I have a dual G4 533 and love the speed this box gives. Under OS X 10.1.5 never a slow down even with running 5 diffrent programs and SETI@Home.



    A Quad system would be "a good thing"® if they could keep all the processors feed with data. I remember Daystar with the Quad system back in the days of Clones. If they could fix the issues for prepress and video pros why can't apple today. They have more software today that would get the speed boost now than any time before.



    Later Steve



    Looking for a Macintosh fast enough to go back in time and trick Bill Gates to say in School!



    [ 07-23-2002: Message edited by: ssmurphy ]</p>
  • Reply 20 of 39
    yonyon Posts: 3member
    [quote]Originally posted by Aphelion:

    <strong>MERSI ME!



    I'd rather have 8 (or even 4) 7410's @ 500 Mhz than 2 74xx @ 1.5 Ghz.</strong><hr></blockquote>



    It's a conceivable alternative for folks

    who:



    -Make extensive use of explicitly MP enabled

    software (video, graphics, and image editing),

    which is able to divide the processing task

    among several processors.



    -Have to process several threads of approximately

    equal CPU consumption concurrently. (Server

    applications, etc)



    Personally I don't do any of this. 2 CPUs

    would be of marginal help, because e.g. the UI

    portion of some audio applications I use

    will schedule separately from the audio

    processing thread (once these applications

    are ported to OSX), but my feeling is that

    it's probably not worth the expense.

    Particularly since the audio is often memory

    throughput rather than processor limited.



    -Yon
Sign In or Register to comment.