Intel wants to get into heavy petascaling
Intel may dominate the list of top supercomputers, but the most intriguing work done in the high performance computing field takes place outside of the Xeon kingdom. Pat Gelsinger, Intel's server chip chief, plans on fixing this problem.
Look at November's Top 500 List, and you'll find 320 64-bit Xeon-based supercomputers. That's 64 per cent of the total systems, although Intel takes even more when you count the 21 Itanium-based systems and 13 running on older 32-bit chips. AMD-based boxes come in a distant second, sitting inside just 16 per cent of the supercomputers.
Intel's wins, however, do little for those who enjoy the most exotic, most powerful machines.
For example, IBM's Power processors still fuel the world's fastest system at Lawrence Livermore National Laboratory. The Blue Gene computers from IBM account for nearly half of the top ten machines in the world.
While not yet officially benchmarked by the Top500 crew, Sun's "Ranger" supercomputer should end up as the second fastest machine on the planet. It runs on AMD's Opteron processors and includes a wide variety of custom networking components made by Sun that Intel's traditional HPC chums HP and Dell have no intention of mimicking.
Looking forward, Los Alamos National Laboratory will unveil a true beast this year - the Roadrunner supercomputer based on Opteron chips. Eventually, this system should use both AMD's processors and Cell chips from IBM, turning Roadrunner into a hybrid dynamo that can handle the toughest visualization tasks.
Jumping a couple more years forward, Sun, IBM and Cray are all working on supercomputer designs that should result in a performance leap that places their hardware one generation ahead of rival boxes. Intel doesn't fit into any of these plans with Sun's system running on its own Rock chips, Cray's running on Opterons and IBM's running on the Power7 chip.
Not so fast, Gelsinger told us, during an interview last week.
"We have a special division focused on HPC - an organization that lives and dies on HPC," he said. "We are not being specific right now about our offerings, but we definitely will be part of the petaflop race.
"We are aiming for the top of the top." Gelsinger declined to provide us with any specifics around an upcoming Intel-based monster. But you could see Intel working with some of the accelerator players to craft a "top" machine that teams co-processors with some upcoming "Nehalem" versions of Xeon. Or maybe Gelsinger is looking even further out at a system which taps the graphics capabilities of the upcoming Larrabee processor.
Intel will release its grand HPC vision later this year at the Intel Developer Forum event in San Francisco, Gelsinger said.
Intel has also enjoyed some HPC success with Itanium, although it seems unlikely that the EPIC chip will lead the company's petascale charge.
Gelsinger declined to say whether or not Itanium has even turned into a profitable venture.
"We don't break out specific financials on any product at that level," he said. "So, I can't answer you, but we continue to make progress. As we gain market share, the financials look better from our perspective."
Itanium followers might find Intel's rhetoric around the Larrabee part amusing. The company has been championing Larrabee's use of the traditional Intel Architecture as one factor that could it help trump more specialized graphics processors and FPGAs for acceleration.
Gelsiinger said that the work needed to write software for new architectures is often measured "in many years - sometimes decades."
While this may be the case, Intel was willing to take the software risk with Itanic. Clearly, the chip taught it a valuable lesson. ®