Intel's MIC processor to power '15 petaflop' supercomputer
Highly parallel MIC chips to enable 15,000 trillion calculations per second
By Patrick Thibodeau | Computerworld US | Published: 10:52, 27 September 2011
Intel's forthcoming MIC processor will be used by the Texas Advanced Computing Center to build a supercomputer with a peak performance of 10 petaflops that will eventually be upgraded to "at least" 15 petaflops.
The system will include a combination of eight-core Intel Xeon chips, which will supply two petaflops of compute capacity, and chips based on the MIC (Many Integrated Core) architecture. The highly parallel MIC processors will provide an additional eight petaflops of performance to the Texas system, code-named "Stampede."
"This is definitely the first serious appearance of MIC in the marketplace," said Steve Conway, an analyst of high-performance computing at research firm IDC.
Related Articles on Techworld
15,000 trillion calculations per second
The Texas Advanced Computing Center, located at the University of Texas at Austin, is "immediately" getting $27.5 million (£17.7 million) from the National Science Foundation (NSF) to build the system, which is expected to be running by January 2013.
The estimated federal investment over a four-year period will be $50 million (£32 million). That includes plans to add future generations of MIC chips, bringing the compute capacity to 15 petaflops, or 15,000 trillion calculations per second. NSF-funded computers are available to scientists to do a wide range of research in areas such as climate, energy, processor improvements and even the spread of diseases.
The supercomputer will also be comprised of several thousand Dell "Zeus" servers.
The MIC chip that the Texas Advanced Computing Center will be using is code-named "Knights Corner," a co-processor designed for highly parallel workloads. It may have more than 50 cores. Knights Corner is competing with Nvidia GPUs - both can be used to help speed up processor power.
Nathan Brookwood, an analyst at Insight 64, said Nvidia is in the catbird seat for people looking for massively parallel types of systems, but the Intel chip has been designed to be amendable to x86 programming environments.
"It's easier to move code over because you do have the x86 compatibility and standard Intel compilers," Brookwood said. But on the other hand, a lot of code that runs in supercomputing environments is adapted to OpenCL, which Intel will support as well, he said.