The Joint Center for Advanced High Performance Computing (JCAHPC) in Japan today released the details of its next generation supercomputer – Oakforest-PACs – which will deliver 25 PFLOPS, use Intel’s Xeon Phi (Knights Landing) manycore processors and Omni-Path Architecture, be built by Fujitsu, and be operational in December 2016.
When fired up, the Oakforest-PACS will be the fastest supercomputer system in Japan for the moment. Twenty-five PFLOPS would have taken the second spot on the TOP500 last November – ahead of Titan (DoE/Oak Ridge NL, 17.5 PFLOPS) but still short of Tianhe-2 (National Supercomputer Center, Guangzhou, 33.8 PFLOPS). There has been speculation that China will field two 100 PFLOPS machines this year, perhaps having one benchmarked in time for the June TOP500.
The details of the system incude:
- “The Oakforest-PACS system has of 8,208 compute nodes, each of which consists of next-generation Intel Xeon Phi processor (code name: Knights Landing), and Intel Omni-Path Architecture (Intel OPA) as a high-performance interconnect, which is a brand-new fabric developed by Intel Corporation. This is the first large scale system with such a processor in Japan. The system is integrated by Fujitsu and its PRIMERGY server is employed as each of compute node. Additionally, the system employs the shared files system (capacity: 26 PB), and the fast file cache system (940 TB), both of which are provided by Data Direct Network (DDN).
- “Peak performance of the Oakforest-PACS is 25 PFLOPS and the total memory capacity is more than 900 TB. All compute nodes and servers of file systems are connected by fat-tree topology based on Intel OPA, which provides full bisection bandwidth. Therefore, flexible and efficient utilization and operation of compute nodes and file systems is available. Moreover, the fast file cache system is equipped with SSD’s and it is suitable for such applications that require higher file I/O performance.”
According to Robert Triendl, SVP at DDN, “The storage environment will consist of multiple shared file system, using DDN’s branch of the Intel Enterprise Edition of Lustre (IEEL) and DDN’s next generation ES14KX Lustre SSU, and providing up to 400 GB/sec of peak Lustre bandwidth. The shared file system will be enhanced by an NVMe-based distributed flash caching layer provided by 25 of DDN’s new IME14KX caching appliance with a total capacity of 940 TB and a peak performance of up to 1.4 TB/sec.”
Fujitsu has taken pains to conserve energy and space use. Oakforest-PACS will be a high-density system with a compact physical footprint, with eight nodes fitting into a 2U rackmount chassis. Advanced “hot water” cooling technology is used to supply cooling water to all the system’s components, each of which have different optimum temperatures. In this way efficient cooling at low power is achieved.
180 PetaFLOP Aurora supercomputer in the USA planned for 2018
SOURCES- Intel, HPCWire, Joint Center for Advanced High Performance Computing