The Taito=
cluster includes a separate partition of compute nodes with dedicated GPU =
accelerator cards. There are 38 of these nodes, and each of them has two NV=
IDIA Tesla K40 GPU accelerator cards installed. These cards are based on th=
e NVIDIA Kepler architecture, which was designed for high efficiency, progr=
ammability and performance. In addition, each compute node hosts two Intel =
Xeon E5-2620-v2 CPUs, based on the Ivy Bridge microarchitecture. Memory wis=
e, there are 32 GB of DDR3 1600 Mhz memory, and the local disks are 500 GB =
SATA3 HDD. Each compute node is connected to the Taito fabric with one Infi=
niband FDR Mellanox ConnectX-3 card. The software stack and operating syste=
m of these compute nodes is similar to the rest of the Taito compute nodes.=

These compute nodes part of a supercomputer from Bull acquired in 2014, = and have been integrated tightly to the Taito cluster to appear as one enti= ty. The system is based on Bullx B700 Direct Liquid Cooled (DLC) Series, th= e computing blades are of model B715 and are fitted into Bullx B700 racks. = There are in total three racks in the system, one for the nodes with NVIDIA= GPUs, one for nodes with Intel Xeon Phi accelerators, and one for the node= s used for administering the system. The network topology is a pruned fat t= ree and is integrated into the Taito fabric.