Original URL: https://www.theregister.com/2011/09/14/appro_tsukuba_xtreme_super/

University of Tsukuba orders 800Tflop Xeon E5 hybrid

Ceepie-Geepie splices x64, GPUs into powerful chimera

By Timothy Prickett Morgan

Posted in Channel, 14th September 2011 15:00 GMT

Supercomputer upstart Appro International has bagged a deal to supply Tsukuba University in Japan with an 800 teraflops machine based on Intel's impending "Sandy Bridge" Xeon E5 processors.

The chip maker is hosting its Intel Developer Forum in San Francisco this week, but is not expected to divulge much about the server variants of the Sandy Bridge design. Nonetheless some details could sneak out here and there.

Appro is not saying much about the feeds and speeds of the system, obviously, given that the processors have not yet been launched by Intel.

The "Frontier" system, as it will be called at Tsukuba, is based on Appro's Xtreme-X design and is being installed at the Center for Computational Sciences. The machine will be a hybrid box consisting of server nodes with two of the Xeon E5 processors and four Nvidia M2090 GPU coprocessors, which made their debut in June of this year.

While Appro is not talking about which Xeon E5 chips the Frontier machine is using, it says that the ceepie-geepie will have 4,288 cores on the x64 side and 548,864 cores on the GPU side, for a combined 802.07 teraflops of aggregate peak number-crunching performance. The GPUs are doing most of the calculating, obviously.

The M2090 GPUs have 512 cores each, so that works out to 1,072 GPU coprocessors in total. At four per node, that works out to 268 server nodes. If you do the math, that means the Frontier machine will be using an eight-core Xeon E5 chip.

The server nodes will be interconnected with ConnectX-3 InfiniBand host bus adapters and InfiniBand QDR switches from Mellanox Technologies. The machine will use the complete Appro HPC software stack, which includes Appro's Cluster Engine cluster management tools. The management nodes will run Red Hat Enterprise Linux, while the compute nodes will run CentOS to save money.

The Frontier system will be delivered to Tsukuba in December and be up and running by the end of January 2012. This is the second Appro system to be sold at the university, by the way. In January 2008, the company won a deal to sell an all-CPU Xtreme-X3 box weighing in at 95 teraflops; the box was installed in May with the assistance of rival supercomputer maker Cray's technical services unit. This particular machine was was based on quad-core Opteron processors from Advanced Micro Devices, with a total of 674 nodes, 2,696 processors, and 10,784 x64 cores.

It wasn't clear at press time if Cray was invited to the Appro party at the University of Tsukuba this time around. The pricing of the clustered system was not divulged. ®