On Sat, May 01, 2010 at 08:45:59PM +0100, Bob W wrote: > > > I've worked in ad agencies that had huge networks of Macs > > -- more than > > > `1000 at BBDO Detroit. Problems were few and far between, non- > > > existent for most users. But you have to have Mac IT > > people. You can't > > > leave it to PC guys. > > > Paul > > > > Thanks Paul. Those PC folk have no idea what can be done. As > > I recall, a few years ago someone at some university built a > > setup of about 300 Mac Pros running as essentially parallel > > processors, and they clocked in faster than any other > > computer or computer array ever, including Crays. Of course, > > such things are only viable for 6 to 12 months when newer > > equipment is introduced that would best them. Such is the > > logarithmic world of the advances in the digital domain. > > 1,000 nodes isn't a huge network.
Hell, 1000 nodes can fit in only a few cabinets. A blade server can get, what, 4 CPUs on a blade? I think you can fit as many as 64 blades in a standard rackmount enclosure. That makes it easy to build a 1024-node system. Of course SGI were building systems with 1024 (and more) MIPS cores a decade ago. Nowadays they use Intel XEON CPUs. Their current fastest system, at NASA Ames, manages roughly half a petaflop from around 6000 nodes (each node has dual quad-core processors). That's with 64 nodes per cabinet. But systems with as many as 8K Processors have been around since before the turn of the century. And these systems have much, much faster interconnects than any setup using looser-coupled Mac Pros (or any other standalone systems). Basically, a system of a mere 300 Mac Pros has never been anywhere pushing the top of the performance charts in the real world; any such claims must have come with so many restrictions and qualifiers as to be meaningless. In 2004 IBM retook the "fastest computer" crown with the first "Blue Gene" system, at a paltry 36 teraflops. (The current incarnation of that architecture has achieved very close to a petaflop). But even 36 teraflops is a bit more than you could get from only 300 Mac Pros. Perhaps you are thinking of this 2008 article: http://www.macobserver.com/tmo/article/The_Fastest_Mac_Compared_to_Todays_Supercomputers/ This compares a single Mac Pro to the LANL Roadrunner system (a top-of-the-line 2008 system). It does point out that the cost per gigaflop of a Mac Pro is significantly lower than that of the Roadrunner system. It also compares the Mac Pro with a 1976 Cray 1. no surprise - the Mac Pro is faster (by about a factor of 1000). More likely, though, is the 2003 Virginia tech system. That managed an impressive 12 teraflops - good enough to be the 3rd fastest system known in 2003. But that used a little more than 300 Power mac G5s - in fact about four times that. Later revisions increased performance by about 25% (and also fixed some of the other, rather more serious, problems), which was good enough to keep the system in the top 10 for the next year, and in the top 20 for another year. -- PDML Pentax-Discuss Mail List [email protected] http://pdml.net/mailman/listinfo/pdml_pdml.net to UNSUBSCRIBE from the PDML, please visit the link directly above and follow the directions.
