HPE Apollo (Hawk)

Image f. HPE Apollo (Hawk)

Hawk is the flagship supercomputer of the High-Performance Computing Center Stuttgart and one of Europe's fastest computing systems. Designed with the needs of HLRS's users in mind, Hawk is optimized to provide large-scale computing power for complex simulations. Its CPU-based architecture also includes a GPU partition that is suitable for high-performance data analytics and artificial intelligence applications, and for hybrid workflows that combine high-performance computing and AI.

With a theoretical peak performance of 26 Petaflops, Hawk debuted in 2020 at #16 on the Top500 List.

Funding

Funding for HLRS's Hawk supercomputer is provided by the Baden-Württemberg Ministry for Science, Research, and the Arts and the German Federal Ministry of Education and Research through the Gauss Centre for Supercomputing.

 Logo MWKLogo BMBF 

System components

System Type: Hewlett Packard Enterprise Apollo

Number of cabinets 44
Number of compute nodes 5,632
System peak performance 26 Petaflops

 

CPU Type: AMD EPYC 7742

CPUs per node 2
Cores per CPU 64
Number of compute cores 720,896
CPU frequency 2.25 GHz
DIMMs in system 90,112
Total system memory ~1.44 PB

 

GPU system type: Apollo 6500 Gen10 Plus

Number of cabinets 4
Nodes per cabinet 6
Nodes in system 24
CPU Type AMD EPYC
GPUs per node 8
GPU type NVIDIA A100
GPUs in system 192
AI performance: node ~5 PFlops
AI performance: system ~120 PFlops
Node to node interconnect Dual Rail InfiniBand HDR200

 

Data storage and networking

Storage: DDN EXAScaler

Disks in system ~2,400
Capacity per disk 14 TB
Total disk storage capacity ~42 PB

If you are a current HLRS system user, click here to report a problem, check on Hawk's operating status, or find technical documentation or other information related to your project.

For Users

Frontend and service nodes

Rack type: frontend and service nodes Adaptive Rack Cooling System (ARCS)
Racks: frontend and service nodes 5 + 2 ARCS Cooling Towers
Frontend nodes 10 x HPE ProLiant DL385 Gen10
Memory of frontend nodes 5 x 1 TB, 4 x 2 TB, 1 x 4 TB
Data mover nodes 4 x HPE ProLiant DL385 Gen10
Service nodes Red Hat Enterprise Linux 8

 

Node to node interconnect: InfiniBand HDR200

Interconnect topology Enhanced 9D-Hypercube
Interconnect bandwidth 200 Gbit / s
Total InfiniBand cables 3,024
Total cable length ~20 km

 

Power and cooling

Power consumption

Maximum power consumption per rack ~90 kW
Power supplies in system 2,112
System power consumption, normal operation ~3.5 MW
System power consumtion, LinPack operation ~4.1 MW

 

Cooling

Cooling distribution units (CDUs) 6
Water inlet temperature (CDUs) 25°C
Water return temperature (CDUs) 35°C
Volume of cooling liquid in the system ~2.5 m³
Water inlet temperature (ARCS cooling towers) 16°C
Water evaporation by wet cooling towers ~9 m³/h

 

Related news

All news

HLRS Achieves ISO 27001 Certification for Information Security Management

On the Way to the Next Generation of High-Performance Computing

Hawk Helps Researchers Identify a New Class of 2D Materials

The Future of Atmospheric Simulation: An Interview with Thomas Schwitalla