Researchgroups in The Netherlands

SURFsara

“Researchers are increasingly confronted with enormous amounts of data (“Big Data”) that must be processed, stored and analyzed. In such research, High Performance Computing (HPC) is an indispensable tool. The Dutch national supercomputer, Cartesius, and the national compute cluster, Lisa, are fast and efficient tools enabling researchers to perform complex computations that are beyond the capacity of “ordinary” computers. As national HPC and e-Science support center, SURFsara focuses on providing a complete package of services in the field of High Performance Computing (HPC), Networking, Data Services, Visualization, e-Science Support and Cloud Services. All services SURFsara offers are free of charge for researchers associated with the Dutch research ecosystem.”

>

We are aware that part of machine learning progress is driven by computational scale as well as by massive amounts of data. In particular for deep learning, it is widely acknowledged that some of the most notable results started appearing after the research community began experimenting with GPU devices for training large neural networks.This enabled both the use of larger networks and faster experiment-to-result cycles, leading to a quicker evaluation and hence evolution of network architectures. However, limits of GPU memory capacity were hit, effectively putting a cap on the network size the “typical” deep learning researcher can work with.
Cartesius, the national supercomputer, features 128 high-end GPUs that machine learning experts can use to either run multiple training experiments in parallel or to experiment with very large networks that would otherwise not fit a single GPU. Besides access to computing facilities, we can offer a tailored program to guide researchers in optimizing and parallelizing the required software. Our consultants are from different scientific disciplines themselves and know exactly how to help researchers benefit optimally from our supercomputer and compute clusters. We believe in the future more and more researchers from the Machine Learning community will start using HPC systems in order to remain competitive, and this motivates us to expand our development expertise in this application domain.
Contact persons:

Valeriu Codreanu, Cedric Nugteren, Walter Lioen