The fifth ranked server manufacturer worldwide–Fujitsu Ltd.–has begun refocusing its high-performance computing (HPC) aspirations on the Xeon Phi coprocessor. Now available for its popular Primergy servers, Intel’s Xeon Phi coprocessor is changing the rules of the game for HPC, according to a Fujitsu, which claims that old-school HPCs with thousands of traditional CPUs consume too much power and makes supercomputers too expensive.
Fujitsu is adopting the Xeon Phi for its high-performance computing (HPC) efforts, incorporating the coprocessor into its Primergy server lines. SOURCE: Fujitsu
What a difference a year can make! Last year Fujitsu was focused on its K-Computer. With 88,128 traditional Sparc CPUs–each with eight cores for 705,028 total, K-Computer was declared the world’s fastest supercomputer in the 2011 Top500 Supercomputer list. In 2012, however, the K-Computer lost out to IBM’s Sequoia, which boasted more than twice as many PowerPC cores –1.5 million. Despite its high speed, the K-Computer consumed almost twice as much power per floating-point operation (825 GFlop/kWatt) compared to supercomputers based on the Xeon Phi, which are rated almost three times higher (2499 GFlop/kWatt) by the Green500 List. In addition, the ability to add 60-cores at a time with a single PCIe card drastically cuts the cost of HPCs.
Fujitsu’s strategy is to adopt the many-integrated core (MIC) model to accelerate all parallel software–as opposed to the graphic-processing-unit (GPU) model which accelerates only single-instruction-multiple-data (SIM) programs. Each Xeon Phi coprocessor is paired with a Xeon E3, E5 or E7 main processor in the server, which Fujitsu recommends for any parallel programs using 100 or more threads (current Xeon Phi coprocessors have 60 cores and 240 threads, but future models could have as many as 64-cores with 256 threads).
Fujitsu’s Primergy models can be configured as blade-, rack-, or tower-servers using traditional Xeon CPUs to supervise its Xeon Phi coprocessors. Rack servers range in size from single-socket models each holding a Xeon E3 main processor with 32-Gbytes of memory and three PCIe slots for Xeon Phi coprocessors, up to eight-socket models holding Xeon E7s–each with 10 cores for 80 main processors per server–with up to 4096-Gbytes of memory and 14 PCIe slots for Xeon Phi coprocessors for a total of 920 x86 cores per server. Larger HPCs can use Fujitsu’s Smart Scale-Out building blocks to configure gigantic computing server farms.