It was almost exactly one year ago when Cray announced it was lowering its 2007 revenue projections after it learned that AMD would not deliver its quad-core Opteron ‘Budapest’ processor on schedule. Budapest is the one-socket, 1300 series Opteron that will inhabit workstations, low-end servers, and Cray supercomputers. To Cray, the chip delay meant that its delivery of quad-core equipped XT4 supercomputers for ORNL and other customers would be pushed into 2008, killing the company’s chances for a profitable 2007. Little did anyone know at the time that the Budapest slip was just a prelude to the larger Opteron fiasco that would play out over the next six months.
Cray had first dibs on the early Budapest processors, but not in time for any 2007 revenue. It wasn’t until May 2008 that the supercomputer maker announced it had completed the upgrade to the Oak Ridge National Laboratory ‘Jaguar’ super with the new quad-core parts. The Budapest chips more than doubled the performance of the ORNL system to a total of 260 teraflops.
But it wasn’t until today that AMD finally announced the introduction of the Budapest chips into general circulation — eight months after the ill-fated Barcelona was launched. A TLB bug discovered toward the end of 2007 forced a relaunch of the multi-socket Barcelonas early in 2008 and undoubtedly pushed out the timeline for their single-socket brethren.
AMD might have given more priority to the Budapest schedule, but that really didn’t play to the company’s strength. In general, AMD is less competitive in the single-socket server space than it is in the multi-socket one, since the advantages of its Direct Connect Architecture are less pronounced for single processor systems. According to IDC, in Q108 Opterons inhabited 14.3 percent of all single-socket x86 servers (23.7 percent in the U.S.). The remainder are hosting Intel parts.
AMD is planning to introduce the single-socket quad-core ‘Suzuka’ as the 45nm replacement for 65nm Budapest by the second quarter of 2009. At this time, AMD has no roadmap to go beyond four cores on a single-socket Opteron platform. This leaves AMD’s post-2010 future with Cray in doubt, especially considering Cray and Intel are now in cahoots.
One last data point. The former HPC marketing director at AMD, David Rich, resurfaced today at Interactive Supercomputing, where he will serve as vice president of marketing. At AMD, Rich developed the company’s cluster market strategy and helped to secure some big HPC wins at Sandia National Laboratories (Red Storm) and the Shanghai Supercomputer Center (Dawning 4000A). Along with CTO Phil Hester’s departure from AMD in April, the company seems less and less committed to the high performance computing market than ever.