Lisa Compute Cluster expanded with GPUs for machine learning
20 JUN 2018
SURF's Lisa Compute Cluster has been expanded with a GPU cluster for extremely rapid computations. These graphics processing units are especially suitable for machine learning, and can dramatically speed up scientific research.
The Lisa Compute Cluster is a centrally managed Linux cluster that is ideally suited for large-scale computations. The new GPU cluster contains 23 nodes, each with four single precision Nvidia 1080Ti GPUs. This equates to an extension of the processing power of 1 Pflop/s (1 petaflop = 1015 single precision floating point computations per second). In addition, each node contains 1.5 TB of high-speed local storage and 256 GB of working memory. There are also two additional nodes, each with four NVIDIA Titan V (Volta) cards, which are intended for testing and experimentation.
Large-scale parallel computations
A graphics processing unit (GPU) is a processor used for large-scale parallel computations. This allows it to take over these tasks from the CPU (central processing unit). In high-performance computing, GPUs are primarily used for machine learning. They are also used for modelling and pattern recognition within research fields such as molecular dynamics, radio astronomy and transmission electron cryomicroscopy.
Walter Lioen, Unit Manager Compute Services at SURFsara:
“We receive a great deal of requests from researchers for services that will enable them to speed up their machine learning algorithms. The new GPU cluster is ideal for this, as a GPU can perform thousands of identical tasks at once. In addition, these GPUs can quickly exchange data with each other. This is useful for large, complex models that work together on the same computation."
Boy Menist, Head of ICT at the Faculty of Natural Sciences, Mathematics and Computer Sciences at the University of Amsterdam:
“The new cluster allows our researchers to train their models much more quickly. One of our PhD students, who is researching how the computer can help with medical diagnoses, managed to complete two papers in one month.”
Lisa's GPU extension will grow following the demand of our user communities. The extra capacity is available for all research universities with an RCCS contract. Pay-per-use is also an option. Research Capacity Computing Service (RCCS) is the SURF insourcing service unit that directly connects computing systems to the SURF member institutions.