GPGPU hardware for Numerical Simulations
Posted: 9 Oct 2013 | 11:47
The NAIS project (Numerical Algorithms and Intellegent Software), which EPCC is a member of, has recently purchased 8 NVidia K20 GPGPUs, and associated computer nodes to house them, for use by NAIS members and researchers.
The GPGPUs complement similar hardware that EPCC already hosts for NAIS (Nvidia Tesla GPGPUs) and will allow NAIS researchers to explore the performance and programmability issues associated with using such computing resources for scientific simulation.
Figure 1: The two nodes installed, over 10 TFlop/s in a 4U compartment!
We are currently busy installing these GPGPUs in two compute nodes that will be attached to EPCC's existing hydra cluster, enabling access to the compilers and libraries installed on that system and facilitating scheduling of access to these GPGPUs through the common batch system used on hydra. Four GPGPUs are each housed in a 2U server chassis, along with two 8-core Intel Xeon processors and 128GB of memory. This provides a double precision peak performance per node of over 5.3 TFlop/s, with 16 processor cores and 7488 GPU cores available in each node.
NAIS is a collaboration between the universities of Edinburgh (including EPCC, Informatics, and Mathematics), Strathclyde, and Heriot Watt. It is funded by EPSRC and the Scottish Funding Council (SFC). These GPGPU resources are funded directly by a grant from SFC to provide high performance computing resources for researchers.
Figure 2: One of the two K20 nodes being built. Each of the silver rectangles is a GPGPU card.
Whilst these GPGPU resources have been purchased for NAIS they are a resource that can be used by the computational simulation, mathematics, and HPC communities in general, so if you are interested in having access to these systems please don't hesistate to contact us.