| PLEX86 | ||
IBM 610 workstation computer 3465IBM 610 workstation computer 3466 You may be tunnel visioning; memory isn't the only wire that has to be going to all CPUs; you also have... +--------------- Only up to a hundred thousand processors, or so. ;-) Seriously, PowerPC (Power4-5 etc.) does 8-way SMP as a basic building block. Intel x86 is at least 4-way and AMD Opterons up to 8-way without bridges, though NUMA not SMP. Opteron's NUMA is coherent, though. +--------------- Opterons are ccNUMA (cache-coherent NUMA, the "cc" is important!), with all of the main memory visible to all of the CPUs, which *is* "SMP" as far as user programs can tell except that a cache-line miss will taken less time if the main memory it's fetching from is "here" rather than "over there". Though Opteron's simulation of a snoopy bus by using broadcast messages starts to slow things down above 4 CPUs or so but can be sped up again by hardware buttist like the "Horus" chip from Newisis.. SGI's Origin & Altix are also ccNUMA, but use a different hardware cache-coherency protocol ("directory-based"), and because of that they scale up to *hundreds* of CPUs max of 512 CPUs per Linux image on the Altix 3700 with *terabytes*(!) of main RAM visible to all of Again, the only difference a user program sees between a "true SMP" and a "ccNUMA" system is that there's more variance in the cache-line miss delays on the ccNUMA system at the same number of CPUs, and that ccNUMA systems can scale up to *far* more CPUs than anyone has even gotten "true SMP" to ever work. IBM 610 workstation computer 3467 I've oft claimed that the non-cache coherency for 801 (from the mid-70s) between the I and D caches for the same processor ... in addition to no provisions... This is not particularly new stuff; SGI was shipping 64- to 256-CPU Origin systems in the late 1990's. -Rob ----- San Mateo, CA 94403 (650)572-2607
|
||||
IBM 610 workstation computer 3466 Alt Folklore Computers from Newsgroups The #1 Usenet Provider on the Internet
|
||||