EPCC hosts its first GPU hackathon
Posted: 11 May 2021 | 10:36
EPCC held its first GPU hackathon this April in partnership with NVIDIA, hosting 28 participants and 20 mentors across seven teams. The event was held virtually due to the ongoing Covid-19 pandemic. Using Zoom and Slack, individual teams were able to work alongside mentors in separate breakout rooms and channels.
Team members originated from universities, companies and HPC centres across the UK, as well as the rest of Europe. The projects they provided involved a range of fields including computational fluid dynamics, machine learning, solid-state physics, molecular dynamics, astrophysics and materials science. Teams developed and programmed their codes on EPCC's Cirrus platform, which has 144 NVIDIA Volta V100 GPUs available.
Each team had different goals for the hackathon, involving parallelisation from scratch, porting from CPU to GPU, scaling up to larger numbers of GPUs or optimising existing GPU codes. The languages and frameworks used varied from FORTRAN with OpenACC to higher-level frameworks like PyTorch. NVIDIA engineers were on-hand providing Q&A sessions during the hackathon (in particular for tools like the NSIGHT profiler), as well EPCC staff to provide support for any hardware or software issues that arose.
By the end of the hackathon, several teams reported they were able to achieve between 2x to more than 8x speedups over original versions of their codes.
Larisa Stoltzfus, EPCC