Even as Silicon Graphics trumpeted on Tuesday a new speed record with the Columbia supercomputer it built for NASA, ZDNet UK sister site CNET News.com has learned, it quietly submitted another, faster result: 51.9 trillion calculations per second.
During the unveiling of the Columbia supercomputer, SGI touted a speed of 42.7 trillion calculations per second, or 42.7 teraflops. That handily beat the machine at the top of a list of the world's 500 fastest machines, NEC's Earth Simulator at 35.9 teraflops, as well as a top challenger, IBM's Blue Gene/L at 36.0 teraflops.
The 42.7 teraflops speed used only 16 of Columbia's 20 servers. That means that 2,048 of the 10,240 Itanium processors in the supercomputer weren't being used -- and the unused chips are the newest generation of Itaniums, each with 9MB of high-speed cache memory, SGI Chief Executive Bob Bishop said.
SGI clocked the full 20-server system at a sustained speed of 51.9 teraflops, according to a source familiar with the test. On a secondary but still scrutinised measurement, peak speed, Columbia ran at 61.0 teraflops, a smidgen ahead of the 60-teraflop speed Intel President Paul Otellini predicted in September.
Full results in the closely watched competition are released every June and November at supercomputing conferences; the newest Top500 list will be released Nov. 8, organiser Jack Dongarra said. Despite the interest in the list, its organisers and others recognise that the speed test, called Linpack, is a convenient but incomplete performance measurement.
SGI's higher speed gives it more of a lead over IBM, but Big Blue's score was performed on 16 September, leaving plenty of time for upgrades and new speed tests.
SGI declined to confirm or deny the statistics, and Dongarra declined to comment.
