It comprises 49 PowerEdge C6420 nodes (with latest-generation Xeon processors providing a total of 2352 cores and 27 PowerEdge C4140 nodes (incorporating Nvidia Tesla V100 GPUs providing a total of 552,960 CUDA cores and 69,120 Tensor cores), plus 700TB of Dell's Lustre storage.
This gives the system a theoretical peak performance (Rpeak) of one petaFLOPS (10^15 floating point operations per second), an Rmax (Linpack benchmark) score of 700 teraFLOPS. Each of its 108 V100 GPUs is rated at 125 teraFLOPS of deep learning performance.
Artemis 3 will support projects in fields including geophysics, cosmology, genomics, proteomics, economics, transport logistics, robotics and medical imaging.
The Dell/Nvidia combination was chosen after a competitive selection process, said Hammond.
Dell EMC high-performance computing lead Andrew Underwood noted that Artemis 3 was purchased in November 2017 and operational by January 2018. The quick deployment was due in part to the pre-engineering and optimisation work performed at the factory, leaving less to be done on-site.
It reached 80% utilisation within two months, with more than 1800 users from 800 projects.
Artemis 3 was connected to its predecessors, forming a single pool of resources, he said. The software used automatically allocates jobs to the most appropriate nodes.
Some workloads run optimally on Artemis 1 or Artemis 2 nodes, Hammond said. For example, Artemis 2 incorporates some large-memory nodes that are still world class.
There are plans to repurpose older GPU nodes for visualisation rather than computation, and in general, the design of the system makes it possible to add nodes containing new technology and to retire old nodes when they are no longer useful.
“The addition of advanced deep-learning capabilities to our Artemis supercomputer — thanks to the Dell EMC PowerEdge C4140 — is a mission-critical dimension of our research infrastructure,” said professor of astrophysics Geraint Lewis, who is also deputy director of the university's Sydney Informatics Hub.
“With a greater number of research problems being data-driven, or more accessible because there is data, our researchers will be able to investigate questions that were previously unanswerable.”
Another consideration is that Artemis 3 makes it easier for the university to build a system that is suitable for anyone from undergraduates to professors. “By exposing our undergraduate and postgraduate students to supercomputing, we provide them with the skills to support and shape the knowledge-based economy essential for Australia's technological future,” said Hammond.