Installation of an OpenPOWER prototype at IDRIS

The first part of the OpenPOWER prototype, Ouessant, arrived at the beginning of February, 2016. It consists of three IBM Firestone compute nodes based on the IBM POWER 8 processors associated with NVIDIA K80 GPUs.

The second part of the prototype, installed in mid-September 2016, is composed of 12 IBM Minsky compute nodes with each containing:

  • 2 POWER8+ sockets, 10 cores, 8 threads per core (or 160 threads par node)
  • 128 GB of DDR4 memory (bandwidth > 9 GB/s per core)
  • 4 Nvidia new generation Pascal P100 GPUs, 16 GB of HBM2 memory
  • 4 NVLink interconnects (40GB/s of bi-directional bandwidth per interconnect); each GPU card is connected to a CPU with 2 NVLink interconnects and another GPU with 2 interconnects remaining (see figure below)
  • A Mellanox EDR IB CAPI interconnexion network (1 interconnect per node)

The tests on these prototypes aim to quantify the work necessary for porting the applications on these architectures as well as the performance obtained. The programming models based on directives (OpenMP 4.n or OpenACC) are preferred and their maturity and performance will be evaluated.

At first, these platforms will be open to the porters of twenty selected codes, representative of different scientific domains.

Later on, they will be open to any community wishing to make tests on these platforms.

The IBM, Nvidia and Mellanox teams will provide user support and will participate in the workshops organised by this technology watch group.