This year the University of Melbourne in partnership with Deakin University, La Trobe University, RMIT and St Vincents Hospital launched a brand-new General-Purpose Graphics Processing Unit (GPGPU) cluster as part of a new high-end compute service hosted by the University of Melbourne. 

Funded through a combination of ARC-LIEF, the University of Melbourne and our University partners, it is operated on behalf of the partnership by University departments of Melbourne School of Engineering (MSE), Melbourne Bioinformatics and Research Platform Services, the service came online in July.  Additionally, MSE contributed departmental funds to augment the service.  The cluster consists of 72 nodes, each with four NVIDIA P100 graphics cards, which can provide a theoretical maximum of around 900 teraflops.  

GPGPUs are a valuable resource for computational science as each GPGPU chip contains thousands of cores that are optimised for certain kinds of tasks.  For example, if the four core CPU in a typical laptop computer were able to compute the flow of particles through a blood ventricle in a certain time, a GPGPU chip might be able to do the same task tens or hundreds of times faster.  This is an example of Computational Fluid Dynamics (CFD), a type of process that is very well suited to the many cores in a GPGPU.  Another research domain well suited to GPGPUs is Molecular Dynamics (MD), where the configurations and interactions of complex molecules and molecular chains are simulated in the GPGPU.

Perhaps the most rapidly expanding area of research to take advantage of GPGPUs is Deep Learning.  Deep learning is a subfield of Machine Learning, where a computer is trained to recognise and identify patterns in sets of data.  For example, a computer can be trained to recognise a cat that it has never seen, by looking at many pictures of cats.  The more pictures of cats used for the training process, the better chance of the machine to identify a new cat.   This becomes very important when you consider that self-driving cars will need to be able to identify not only cats, but all sorts of things in all sorts of configurations, in real-time.

The new GPGPU service has seen a rapid take up, reaching full capacity (usage >95%) within six weeks of launch. Coupled with high-performance storage, the service is already supporting almost 100 research groups across the five partners and has processed over 100,000 jobs.

Research Platform Services has started running training to assist researchers to prepare their jobs for the GPGPU environment, with more courses including GPGPU programming planned for 2019.

While not every computing workload is well suited to GPGPUs, more and more applications are including modules specifically for GPGPU (Matlab, anyone?).  If you think you might have a computation challenge that might benefit from a GPGPU, or would just like to know more about them, please email hpc-support@unimelb.edu.au and we will be in touch.