Cray plans to create a new supercomputing platform combining four different types of processing capability in a blade server architecture. The platform will run Linux on Advanced Micro Devices's Opteron processors, and should be on the market by 2010, a spokesman said Monday.
The company anticipates changes in the nature of high-performance computing applications: instead of being composed purely of scalar calculations, or entirely of vector operations, they will increasingly be a mix of the two. Cray intends that its future supercomputers match that mix.
"Even today, you see heavily vectorized applications where 1 percent is scalar, but if you don't have a scalar processor for that 1 percent, it might take 10 times as long to run the application. You could just fall off a performance cliff," Cray spokesman Steve Conway said.
The first phase of the transformation is to offer a single programming environment across the different types of supercomputer Cray makes. That environment will be based on Linux, and will start to appear on the market in 2007, Conway said.
The next phase will see the introduction of a single system capable of integrating different blades optimized for either scalar processing, vector processing or multithreading. The blades could also contain hardware accelerators dedicated to particular functions. Compilers will analyze code and target it to the most appropriate type of blade for the task.
"Cray has a big head start because it has a compiler that can look at code and say, 'This will run best on a scalar processor, and this will run on a vector processor,'" Conway said.
All the blade types will contain Opteron processors, but some will contain additional elements integrated on the same piece of silicon, dedicated to tasks such as vector processing, he said. The additional functions will be incorporated in the form of ASICs (application-specific integrated circuits). The silicon will most likely be manufactured by Cray's current ASIC partner, Texas Instruments, Conway said.
Cray also will develop systems that can automatically allocate the different kinds of computing resources to an application. That phase of development depends on Cray's bid for funding from the Defense Advanced Research Projects Agency (DARPA) High Productivity Computing Systems program, and would have to be delivered by 2010, Conway said.
Sun Microsystems and IBM are competing with Cray for the same DARPA funding, which could be worth as much as US$200 million, he said.
IBM is already experimenting with combining scalar and vector processing in a single system. On several occasions in recent months, it has demonstrated a server using blades based on its Cell multicore microprocessor. Cell contains a central core based on the PowerPC scalar architecture, and eight additional vector processing cores.
Learning to program such hybrid parallel systems efficiently will be a challenge. Languages such as Universal Parallel C and Co-Array Fortran already exist to make this easier, but each of the bidders for the DARPA funding is developing its own language. Cray's, developed with the California Institute of Technology, is called Chapel.