An enhanced online graphics processing unit facility for Early Career Researchers

Lead Research Organisation: University of Kent
Department Name: Sch of Biosciences


To identify current needs of ECRs and decide on the best way to use this funding allocation to further their research, a consultation was held between current and recent eligible EPSRC grant holding ECRs from schools across the science faculty. In consultation with other ECRs and Directors of Research within the Science faculty, they identified a need for a centrally managed high-power parallel GPU computing facility. This would align well with internal strategies for equipment within the EPSRC disciple relevant schools, and would add value to and complement the existing investment within the university.
With this bottom up approach we are confident that the proposed purchase of a centrally managed high-end multimode GPU server will provide maximum benefit to ECRs cross-faculty and enable their endeavours in world class cross-discipline research. From the perspective of scientific research, including large-scale data analysis, modelling and visualisation, the most important practical reason for using GPUs is that they enable rapid model prototyping and evaluation. These procedures, if performed on GPUs, will significantly reduce the barrier to building complex models from large scientific datasets, and for seeding potential applications of this research. This resource will hence accelerate advances across the breadth of scientific challenges being pursued by ECRs at Kent.

Planned Impact

The recent upsurge in interest in Artificial Intelligence is driven by its potential for transforming diverse fields ranging from policing, transportation and healthcare. The central development in this field of research that has generated this interest is the maturation of deep learning with artificial neural networks, and its confluence with the emergence of very large training datasets, and significant technological advances in the availability of specialised parallel computing hardware. In particular, modern software packages for deep learning, e.g., Google TensorFlow, rely centrally on matrix computations for simulating neural networks. These computations are inefficient on conventional central processing units (CPUs), and are best performed in parallel on Graphics Processing Units (GPUs), which were expressly designed for fast, parallel calculations on matrices. For example, deep learning applications like automatic image and speech recognition can be 15-33 times faster when implemented on GPUs. Further, GPUs have a large memory bandwidth to deal with the data flow involved in these computations. From the perspective of scientific computing, including large-scale data analysis, modelling and visualisation, the most important practical reason for using GPUs is that they enable rapid model prototyping and evaluation. The purchase of a high power GPU server will facilitate many different areas of research that can make use of cutting-edge parallel processing unit within modern GPUs. The proposed centrally managed high-power parallel GPU computing facility will have major impact on the productivity of ongoing and future research of ERCs across the science disciplines within the EPSRC remit. This investment aligns closely with the University of Kent institutional plan and the faculty research strategy for investing in world class facilities for the widest range of cross-disciplinary research. In addition, this proposed investment fits well with EPSRC world class strategy.


10 25 50