Reconfigurable Architectures for Scientific Computing

Supercomputers play an important role in the field of computational science, and are used for a wide range of computationally intensive tasks in various fields across the engineering and sciences spectrum [1]. As the computational power of these systems continues to grow, there is typically a hard constraint on the maximum power that can be drawn by a system. It is therefore vitally important that new systems are procured with energy-efficiency in mind, and that applications are developed to achieve the maximum performance while making every effort to reduce their power requirements [2-4].

Although the theoretical peak performance of the fastest machines is rapidly approaching the era of Exascale computation, real applications are typically realising only a fraction of this – representing a significant waste of resources. One potential solution to this is the use of reconfigurable architectures — architectures that can be precisely configured at runtime to contain only the computational units required, in a configuration which is more tailored to the needs of the target application [5,6]. 

This research project seeks to investigate the applicability of reconfigurable architectures (e.g. field-programmable gate arrays [FPGAs]) to the computational sciences.

Specifically, the objectives of this project will be to (i) evaluate the programmability of FPGAs for applications with a heavy reliance on floating-point calculations; (ii) port a small number of proxy applications from the physics domain to FPGAs; and finally, (iii) analyse the performance of these applications in terms of their runtime and energy performance, in order to quantify the trade-offs that may be available.

Familiarity with parallel programming and programming FPGAs is desired.

Research areas: High-Performance Computing, Embedded Computing, FPGA, Energy-aware computing

For more details please contact Dr. Steven Wright

E-mail: [Email Address Removed]