Computational Modelling Group

New Facility for Physics and Astronomy

Physics and Astronomy(P&A) have purchased a new compute server to support computer modelling across their Astronomy, Theory and Quantum Light & Matter groups. The groups need to quickly run and debug their own mix of modelling code including CPU and memory intensive computation and GPU based machine learning.

Models typically require large amounts of system memory (>500GB), GPU memory (>32GB) and graphical outputs, meaning dedicated resources are a huge benefit rather than having to queue workloads on larger command line only systems. To deliver this P&A have invested in a single high spec machine available to their staff and researchers, featuring: 128 AMD CPU threads 2TB RAM (offering multiple GBs per CPU thread) 20TB Fast local storage 120TB of network attached storage. NVIDIA Ampere A100 80GB PCIe Gen4 GPU

Delivering this project has involved collaboration between Physics and Astronomy, Faculty of Engineering and Physical Sciences and iSolutions, and will be of immediate benefits in a number of areas.

The Theory group will use the new compute server for calculations in Quantum Field Theory. In particular it will be used to compute scattering amplitudes, which are quantum probabilities describing how elementary particles interact with each other. The large memory available in the machine will be instrumental in solving linear algebra problems that are required to produce state-of-the-art results, using novel techniques developed at Southampton. Another important area is extracting physics results from correlation functions produced on supercomputers (eg STFC DiRAC Extreme Scaling systems and others in Europe, US and Japan). Many correlators need to be analysed together, to improve the isolation of the desired particle states, to incorporate small corrections from electromagnetism or isospin-symmetry breaking by different quark flavours, or simply because of the complexity of the combinations needed for the quantities of interest. Codes are run hundreds of times as they are developed, tested and reworked to establish final analysis strategies and produce results with full statistical and systematic errors.

The Astronomy group will use the capabilities of the new compute server to reduce and store data from various large sky surveys that have an ever increasing demand on disc space and parallel processing power.

The Quantum Light & Matter group will use the new compute server for machine learning where a convolution neural network is employed to solve the phase retrieval problem. Phase retrieval is the process of recovering images and phase information from Fourier amplitude measurements of a speckle pattern. Recovery of images from high resolution 3D speckle pattern data requires dedicated GPU hardware with relatively large memory for data storage and processing.