Deucalion

Deucalion featured image

System Overview

  • Name: Deucalion
  • Location: Portugal
  • Owner: FCT
  • Operator: CNCA
  • Released in: 2023
  • Primary focus area: Industrial

Architecture and Performance

Deucalion (theoretical maximum of 10PFlops) has three main partitions: ARM, x86 and GPU-accelerated. The ARM partition is the largest with 1632 nodes equipped with Fujitsu’s A64FX chips, with maximum LINPACK performance of 3.96 PFlops (theoretical peak is 5.01 PFlops) and excellent energy efficiency, making the top 100 of the Green500 list. The regular x86 partition (measured at 1.86 PFlops) is composed of 500 nodes, each with 2x AMD EPYC 7742. The GPU-accelerated partition has 33 nodes equipped with 4 Nvidia A100 each (both the 40GB and 80GB vRAM versions) coupled to the same AMD CPUs.

For the most up-to-date information visit Deucalion website.

Achievements and Impact

Rankings & Records:

The ARM partition is currently on position 257 of the Top500 list and at 99 of the Green500 list, due to the excellent energy efficiency of the A64FX chips. The smaller x86 partition boasts a LINPACK performance of 1.86 PFlops.

Scientific Contributions:

The GPU partition is a starting point for larger AI workflows, including the first Portuguese LLM. We are also starting up the Fujitsu’s optimised quantum simulator qulacs, able to use the A64FX’s capabilities to their fullest for a maximum 40 qubits using 1024 computing nodes. Our ARM partition can also kickstart specific SVE vector code development and porting.

Sustainability and Future Outlook

Environmental Impact:
Use of energy efficient ARM architecture for typical HPC workloads
Use of efficient cooling infrastructure, boasting a 1.2 PUE (Power Usage Effectiveness)

Future Developments:
Use of ARM chips for AI/LLM workloads (primarily inference)
Quantum simulation using ARM chips

Contact Information