CEA
HPC High performance computing
HomeComputing centersSoftware for HPCResearch and developmentCollaborations
Computing centers / TGCC / Joliot Curie

TGCC Joliot Curie

Logo GenciLogo TGCC


 
The most powerful French supercomputer

dedicated to French and European research

This new 9,4 petaflops Joliot-Curie supercomputer is in production in CEA’s Very Large Computing Centre (TGCC) in Bruyères-le-Châtel, France, since mid-2018.

Joliot Curie is a BULL Sequana X1000 supercomputer, funding by Genci, and have 2 compute partitions:


Joliot-Curie

Main partitions

The first phase of the supercomputer is composed of Bull Sequana X1000 cells with a peak power of ~ 9 Pflops composed by the following partitions :

  • SKL Irene (Skylake):
    • 1,656 dual-processor Intel Skylake 8168 fine nodes at 2.7 GHz with 24 cores per processor, for a total of 79,488 computing cores and a power of 6.86 Pflop/s,
    • 192 GB of DDR4 memory/node,
    • Infiniband EDR interconnect network.
  • KNL Irene (Knight Landing):
    • 828 Intel KNL 7250 manycore nodes at 1.4 GHz with 68 cores per processor, for a total of 56,304 cores and a power of 2 Pflop/s,
    • 96 GB of DDR4 memory + 16 GB of MCDRAM memory/node,
    • Bull eXascale Interconnect network (BXI).

    Since the end of 2019, a second phase composed of Bull Sequana XH2000 cells adds the following partitions :

  • AMD Irene ROME
    • 2292 dual-processor AMD Rome (Epyc) compute nodes at 2.6 GHz with 64 cores per processor, for a total of 293,376 computing cores and a power of 11.75 Pflop/s,
    • 256 GB DDR4 memory/node,
    • Infiniband HDR100 interconnect network.
  • Irene V100 - data processing/IA
    • 32 hybrid nodes with 2 Intel CascadeLake 20-cores 2.1 GHz processors and 4 NVIDIA V100
      GPUs per node, for a total of 128 GPUs and a power of 1.13 PFlop/s.

    As part of the Fenix project, the following partitions are also available :

  • Irene V100L
    • 30 hybrid nodes with 2 Intel CascadeLake 18-cores 2.6 GHz processors and 1 NVIDIA V100 GPU per node,
  • Irene V100XL
    • 2 hybrid nodes with 4 Intel CascadeLake 18-cores 2.6 GHz processors, 3 To memory and 1
      NVIDIA V100 GPU per node.
    Common resources :
    • Irene hybrid : 20 vizualisation nodes with NVIDIA-P100 co-processors,
    • Irene xlarge : 5 so-called big memory/nodes for pre/post processing with 3 TB of memory/node and an NVIDIA-P100 co-processor,
    • A local storage, Scratch type, of 5 PB with a bandwidth of 300 GB/s.
    Exploratory partition - ARM :
    (available in the first half of 2021)

     


    flbJoliot-Curie in video

    Joliot-Curie June 2019 (in french)
     
    Street art with C215 in a computing room, a first world artistic performance (in french).