The Big Mac Saga continues! To recap, Virginia Tech's cluster of 1,100 dual-processor Power Mac G5s came online for performance testing at the beginning of the month, and on the 15th, WIRED quoted a Mr. Jack Dongarra ("one of the compilers of a Top 500 list") as saying that "early benchmarks" from a test of 128 of Big Mac's 2,200 processors showed performance a whopping "80 percent of the theoretical peak" of 17.6 teraflops. If the cluster could hold that ratio when all 2,200 processors were cranking away, then Big Mac had a decent chance of being ranked as the second most powerful supercomputer on the planet, and thus qualifying for a free Funny Face breakfast at the International House of Pancakes. (Japan's top-ranked Earth Simulator gets the full-on Rooty Tooty Fresh and Fruity®.)
Unfortunately, just two days ago, the New York Times published a followup from Mr. Dongarra, who revealed that testing of the full cluster showed Big Mac's score as "only" 7.41 teraflops-- a mere 42% of the system's theoretical peak, and even marginally slower than a Xeon-based cluster we had expected to see stomped flat by Divine G5 Muscle. At 7.41 teraflops, Big Mac would still be one of the top ten supercomputers in the world, but it sure as shootin' wouldn't rate free breakfast treats at the blessed Eatery with Eight Syrups.
Faced with the surprisingly low score, Virginia Tech stated that it was "still finalizing results" and that Big Mac's final, official score "might be significantly higher." To be honest, we figured that was just spin-- but faithful viewer Karl Kornel informs us that Virginia Tech apparently wasn't kidding. Karl found a MacBidouille article that links to a PDF of an 80-page paper penned by-- you guessed it-- the ever-lovin' Jack Dongarra. (That little fella is everywhere.) This report is dated October 22nd, 2003-- a day after the Times article was published-- and refers to Big Mac's measured performance as being rather higher: 8.164 teraflops.
Over three-quarters of a teraflop improvement in the space of a day or two? Nifty! Unfortunately, Big Mac's latest numbers leave it squarely in fourth place; while the new score does indeed lay the smackdown on that 7.63-teraflop Xeon cluster we'd been worrying about (which has since slipped into seventh place), unfortunately, there's a new player in town. An Itanium 2-based cluster just slid into third with a measured performance of 8.633 teraflops-- and with 176 fewer processors than Big Mac, to boot. D'oh! So much for a clear victory of IBM processor technology over Chipzilla. Granted, the Itanium 2 is strictly a server and workstation chip, while the G5 is available in $1999 personal computers, but a blowout of All Things Intel would have been sweet.
There's still hope, though; while the Itanium 2 cluster has a theoretical peak performance of 11.616 teraflops, Big Mac can theoretically peg out at 16.896-- 17.6 with all 2,200 processors chugging away. (Dongarra's paper lists Big Mac as having only 2,112 active chips, so apparently 44 of those Power Macs were off watching "The Next Joe Millionaire" during the last test). What that means is that Big Mac has much more potential than the Itanium cluster, but is currently operating at just 48% capacity while that pile of Intels is ticking away at over 74%-- so all Virginia Tech needs to do is figure out how to get those Macs to start pulling their weight.
Clearly there's room to grow, and Virginia Tech just needs to boost processor morale. AtAT sources reveal that the recent .75-teraflop increase comes courtesy of a rousing visit to the Big Mac facility by the Hokie cheerleaders, and now it's time to kick things into high gear; may we suggest a motivational screensaver from Successories? Sure, it's Windows-only, so they'd have to install Virtual PC to run it, but c'mon-- the G5s in those Power Macs are apparently 52% idle anyway, so what can it hurt?