The second pilot phase of the European supercomputer, LUMI, using the system’s GPU capacity, is complete. The system’s performance has been verified with accepted benchmarks and, therefore, LUMI is officially ready to serve Danish scientists in its full capacity!
Though LUMI was officially inaugurated in June 2022, and Danish researchers have been able to use the CPU partition (LUMI-C) of the supercomputer since, the GPU partition (LUMI-G) has not been available until now. The availability of LUMI-G is therefore a great news to Danish researchers – especially as the GPU partition is by far the largest part of the supercomputer: LUMI-G consists of 2560 GPU nodes, each node with one 64 core AMD Trento CPU and four AMD MI250X GPUs.
Associate Prof. at the Department of Mathematics and Computer Science at SDU, Benjamin Jäger, is one of the few Danish researchers who has already tried the LUMI GPUs. As Denmark is a member of the LUMI consortium, 3 Danish projects were able to test their code on LUMI-G during the testing phase. Benjamin’s project was one of the these 3 Danish pilot projects (see the full list of pilot projects here).
“In my opinion, LUMI-G work extremely well from the beginning, there were hardly any restarts or crashes, which I expected during the testing/pilot phase. I also was impressed by the available software and the support that the LUMI team provided, in particular to my team members. My team and I were using the GPUs relatively early on. It took us a bit of time to adapt our codes to run on the AMD GPUs (as our code is developed for NVidia GPUs), but it turned out great. One big advantage of the LUMI-G hardware is the larger amount of memory on the GPU. This is very beneficial for our simulations. The lattices we use in our simulations need a substantial amount of memory in order to be able to fit all computations on the GPUs, which gives us the best performance. Unfortunately, the pilot phase is over, we would have loved to continue more,” says Benjamin Jäger.
SDU will make a call for applications for LUMI-G resources
SDU owns a share of the GPU nodes on LUMI and the SDU eScience Center will therefore in the coming weeks announce an internal SDU call for applications for LUMU-G resources. SDU researchers can also apply for the national portion of LUMI-G resources via DeiC, who regularly announce calls twice a year (see the DeiC website).
