Marking a new era in supercomputing, the MilkyWay-2 system today was recognized as the most powerful supercomputer in the world according to top500.org. The system registered a performance of 33.86PF on the industry standard Linpack benchmark on a system with 16000 compute nodes, including over 32000 of upcoming Intel Xeon processor E5-2600 v2 and over 48000 Intel Xeon Phi coprocessors. The total system power consumption was 17.8 MW, which made the system one of the most power efficient systems on the top 500 list with a performance per watt of over 1.9 GF/W.
The system was designed and developed by the National University of Defense Technology (NUDT) and uses what is called a “neo-heterogeneous architecture”. The previous TianHe-1A system was built using a fully heterogeneous architecture combining traditional CPUs with attached accelerators. While this architecture made great strides in energy efficiency, in this architecture, the CPUs and accelerators use a very different programming model, which makes the development of scientific applications more complex and time consuming. In a “neo-heterogeneous architecture” while the hardware architecture has multiple classes of hardware compute capabilities, they are all accessed by a common programming model, streamlining development and optimization processes.
The MilkyWay-2 system is based on a NUDT designed proprietary high-speed network called TH Express-2, and has over 1 Petabytes of memory. It has over 12 Petabytes of global shared parallel storage, and includes 125 cabinets for computation, 13 cabinets for communication and 24 cabinets for storage.
The MilkyWay-2 system is the largest All-Intel based system ever built. It is among the first to use the future Intel Xeon Processor E5-2600 V2 processors, and the largest to-date installation of Intel Xeon Phi coprocessors.
In MilkyWay-2, the proprietary high-speed interconnection network is provided, which adopts the opto-electronic hybrid interconnection and the fat-tree topology. With the bi-directional bandwidth at 160Gbps, the network itself is balanced and scalable.