AMD and HPE unveil support of Adastra supercomputer
GENCI, the French national agency for High-Performance Computing and CINES, the National Computing Center for Higher Education (one of the 3 French national centers), together with Hewlett Packard Enterprise (HPE) and AMD, are proud to unveil the acquisition of “Adastra” supercomputer and its upcoming deployment during next spring.
At 2022 horizon, Adastra will be one of the most powerful supercomputers for academic and industrial open research in Europe, thanks to its state-of-the-art architecture, which will complement GENCI systems already available at the two other national computing centers (TGCC for the CEA and IDRIS for the CNRS-INS2I).
This new leading-edge system “Adastra” takes its name from the Latin sentence “Per aspera ad Astra” meaning “Through hardships to the stars”. Adastra will indeed provide scientists with innovative and large-scale computing capabilities on the difficult paths of research with more than 70 PFlops/s, more than twenty times the computing power of the current system in production at CINES. This will strengthen the position and the means of French research on the exciting path to Exascale.
GENCI has selected the HPE offer based on the HPE Cray EX system after an open competitive dialog that happened during the pandemic period, for 18 months. The selection has been based on a TCO approach, relying both on sustained performance and optimized energy consumption for full-fledged scientific and industrial applications representative of a typical production workload and an innovative process of an integrated collaboration called “contract of progress”, starting for 2 years in November, earlier than the delivery of the system.
Thereby, Adastra will be based on a balanced and modular converged architecture with 2 complementary compute partitions, to answer the various needs of hundreds of daily users:
Partition 1 has manycore scalar nodes, each one based on next-generation AMD EPYC processors, codenamed “Genoa” with 768 GB of DDR5 memory and one 200 Gbps Slingshot 11 NIC
Partition 2 has hybrid nodes, each one with an optimized 3rd Gen AMD EPYC CPU with 256GB of DDR4 memory and four AMD Instinct MI250X OAM accelerators, each with 128 GB HBM2e, for a total of 512 GB GPU memory, and four 200 Gbps Slingshot 11 NICs
The accelerated partition (GPU) is scheduled to be commissioned in the spring of 2022 while the scalar (CPU) partition is planned to be commissioned end of 2022.
The whole system, federated by HPE Slingshot, a high-performance Ethernet fabric designed for HPC and AI solutions, will access a two-level Cray ClusterStor E1000 Lustre parallel file system from HPE providing close to 2PB with a throughput of 1.3 TB/s based on the level 1 full-flash scratch and a second level based on 24 PB @ 250 GB/s based on fast rotative disks.
In the goal of optimizing global energy consumption, the solution provided by HPE is among one of the most efficient by cooling 97% of the heat generated by its machine through warm water liquid cooling associated with a PUE of 1.10 leading to a maximal sustained consumption of 1.59 MW.
Finally, as part of the contract, HPE and AMD are engaged with CINES and end users’ communities in a 2-years contract of progress toward jointly porting, optimizing, and scaling out HPC and AI applications on the accelerated partition of Adastra. These activities will leverage the AMD ROCm open software platform using the HIP and OpenMP programming models, compiler toolchain, and analysis tools, thus participating in the French global effort toward Exascale-enabling of scientific and industrial applications.