Many new and upcoming HPC systems provide computing power through GPUs. Relatively little research software uses such hardware and less still can fully exploit multiple GPUs simultaneously. The Research Software Group (RSG) is in a position to offer time-limited support to researchers at the University of Birmingham with research software that currently runs on GPUs and, given some software development effort, would be able to use or better use multiple GPUs simultaneously, ideally across multiple HPC compute nodes.
We invite expressions of interest from any research groups that are using software that would benefit from this opportunity. We hope to establish collaborative projects starting in either April or October 2025 and lasting up to 6 months, with 45% FTE from one of our Research Software Engineers. The arrangements will depend on the number and nature of potential projects we receive. We aim to collaborate with the researchers to produce research outputs to disseminate the results and experience of extending their code, potentially in specialist research software journals or conferences.
Please express your interest by emailing bear-rsg@contacts.bham.ac.uk by 24 Jan 2025, including “Baskerville Multi-GPU RSE Call 2025” in the subject line. We invite short informal enquiries but the more detail you provide, the more easily we will be able to evaluate the suitability of your project. We are ultimately interested in short answers to at least the following questions (with links where relevant):
- Is the relevant source code available? Is it open source?
- What programming languages or libraries does the code use?
- To what extent is the code able to use multiple GPUs? Can it use one GPU? Multiple GPUs on a single node? Multiple GPUs spread across multiple nodes?
- Has the code already run on a GPU-accelerated HPC system (e.g. Baskerville or BlueBEAR)
- What is the potential wider impact of accelerating the code on multiple GPUs?
- Do you have sufficient data to allow us to run the modified code on a GPU supercomputer several times and get statistically useful performance results?
- Is there already evidence that the program should be able to scale well across multiple GPUs?
These are not requirements but will allow us to more effectively prioritize projects.