Slashdot Mirror

← Back to Stories (view on slashdot.org)

Apple To Discuss HyperTransport For Future Macs

Posted by timothy on from the think-hyper dept.

macrealist writes "CNET is reporting that Apple will discuss the use of HyperTransport in Macs at the Developer's conference. The interesting thing is that the article claims that Apple is not likely to use hypertransport to link the CPU to the memory, but instead to link chipsets together because IBM would have to 'to adapt it to the Power architecture.' But according to arstechnica, the 970 does have a frontside bus that operates at similar speeds to Hypertransport."

2 of 86 comments (clear)

  1. Re:Mac Clusters? by danigiri · · Score: 5, Informative
    Much as we all would love to see it, getting HyperTransport outta-box would involve a lot of tradeoffs that would lower its speed, ending up possibly lower than FW800 (that is designed from the ground up to be external).

    <pedantically> I think that Apple has already developed a tried and true solution for external, non-ethernet-based, high-speed data transfer. It is called FireWire800.

    Of course, an IP substack can be built on top of the FW, to have additional networking options. (Check out)</pedantically>

    0.02â

  2. Re:What about upgrades? by Anonymous Coward · · Score: 5, Informative

    To do it right, you'd have to get a new bus as the chips are being strangled by bandwidth bottlenecks on current bus designs.

    I don't know where these rumors get started.

    The combination of fat caches, low latency, and predictive fetching basically negates the memory bus bottleneck in the current-generation (MaxBus-based) Power Macs. Even in SIMD instances, the processor generally doesn't have to wait on data that much. (This is especially true in SIMD instances, because these are almost always sequential-read applications, which makes those fat caches and predictive fetching work up a sweat.) Consider Apple's AltiVec-optimized BLAST, for instance. It's 10X faster than BLAST on a Pentium 4. It's not memory-bound. It's compute-bound.

    What's that famous Seymour Cray quote? "A supercomputer is a device for turning compute-bound tasks into I/O-bound tasks."

    So if anybody produces PowerPC 970 upgrades with MaxBus interfaces, they're almost certain to be good buys. Unless they cost thousands of dollars, of course. A dual-1.8 GHz (pulled that number out of my ass, guys) Power Mac G5 (pulled that out of my ass, too) will be faster than a dual-1.8 GHz upgrade in a MaxBus G4, but it'll still be considerably faster than the G4 was originally.