The University of Sydney has revealed details of a $2.3 million upgrade to its Artemis high-performance computing (HPC) cluster that it says will significantly boost artificial intelligence research.
Artemis is the university’s flagship HPC facility. It was initially rolled out in March 2015, offering a university-wide HPC capability that is intended to complement national facilities such as the NCI Raijin supercomputer based at ANU in Canberra.
Artemis has been “extremely successful” and demand for the platform has been “fantastic,” said Jeremy Hammond, the university’s director of strategic ventures
The university’s 2016-2020 strategic plan stated that it would make “transformational investments” in its advanced computing infrastructure, including a major expansion of Artemis.
The “next generation” of informatics and big data are underpinned by access to “key computational high-performance resources,” Hammond said. “On top of that we’re seeing an emergence of deep learning and other artificial intelligence techniques and they’re rapidly increasing the demand for high-performance computing, in particular accelerated high-performance computing.”
A previous expansion of Artemis took place in October 2016, when the university boosted the number of cores from 1512 to 4264. Within a “few months” the university’s researchers and industry collaborators were utilising all of the additional capacity, Hammond said.
The next phase of the university’s HPC capability, Artemis 3, doesn’t just add compute cores but also “rapidly expands into a new dimension of HPC around neural networks... and other artificial intelligence methods”.
The $2.3 million Artemis 3 can deliver a theoretical peak performance of 1 petaflop.
The Artemis 3 expansion is based on Dell EMC’s PowerEdge HPC platform and adds 49 PowerEdge C6420 nodes (for a total of 2352 additional x86 compute cores and 9TB of memory), as well as 27 PowerEdge C4140 nodes, each equipped with four Nvidia Tesla V100 GPUs, and a further 700TB of Lustre storage.
(Artemis 2 only offered access to five GPU compute nodes based on PowerEdge R730 servers that were each fitted with two NVIDIA K40 GPUs.)
The density of Artemis 3 will reduce the “power cost per calculation” for researchers, said Dell EMC HPC and AI lead for Asia-Pacific and Japan, Andrew Underwood.
The university purchased Artemis 3 in November. The expanded infrastructure was installed by the end of January.
Two months after the release of the additional capacity, Artemis is already back at 80 per cent utilisation, according to Hammond.