Norwegian researchers are reporting significant performance gains and a smooth user experience during the pilot phase of Olivia, Sigma2's newest supercomputer.
As the pilot phase for the new Olivia supercomputer nears its conclusion, feedback from the initial user groups has been overwhelmingly positive.
Olivia, equipped with cutting-edge NVIDIA GH200 Grace Hopper Superchips, is demonstrating its capability to accelerate research and handle massive datasets more efficiently than previous systems.
Several Norwegian research teams have had the opportunity to test the new machine, and the results are already highlighting the system's potential.
Researchers Report Major Speedups
Børge Arntsen from the Department of Geoscience and Petroleum at NTNU, who was also a pilot user on the LUMI supercomputer, has seen remarkable results with his finite-difference wave equation solvers.
— Olivia works well and performance is good. The speedup relative to multi-core CPU is around 30. Given the number of GPUs, the implication for research is that we can do a lot more in much less time and get results which would have been difficult to obtain with current systems, says Arntsen.
He also notes that the new hardware shows a significant generational improvement, with a single GH200 running his code approximately twice as fast as the previous generation A100 GPUs. For Arntsen, the user experience was also a key highlight.
— The Lumi/AMD system was difficult to use, mainly because of confusing/lacking docs. The Olivia system works without any problems using the default environment.
Other pilot projects echo this sentiment. The team at DigiFarm, a startup developing AI models for agriculture, compared their model training performance directly against the LUMI supercomputer. Their tests showed that training is approximately three times faster on Olivia, with GPU utilisation reaching 90%, compared to 75% on LUMI.
Alexander Hanke, a researcher working with the REEF3D computational fluid dynamics software, is also satisfied.
— For my research, Olivia provides faster time to solution than Betzy. I have not experienced any problem regarding connectivity, uptime or data access.
A Collaborative Pilot Phase
The pilot phase has been a collaborative effort, with researchers and the NRIS user support team working closely together to fine-tune the system. While some users are still in the data preparation phase or working to install custom software packages, the interaction has been productive.
— Communication on the Discord server has been very responsive and fruitful, notes Hanke.
Yiyao Chen, PhD student at the University of Bergen, is preparing to use Olivia for deep-learning training with multi-terabyte geoscience datasets. She appreciates the work done to create a stable environment.
— So far, I have found the provided containers very convenient. It is clear that a lot of work has gone into making the base environment stable and performant, which makes it much easier for new users like me to get started, says Chen.
The pilot period is planned to continue through most of September before the official production start on October 1st. The successful pilot phase indicates that Olivia is well-positioned to become a critical resource for Norwegian researchers, driving innovation across a wide range of scientific fields.
More about the pilot projects
1. Imaging of Seismic Data
Led by Børge Arntsen at the Department of Geoscience and Petroleum at NTNU, this project uses advanced wave equation solvers to create detailed images of the Earth's interior from seismic data. The immense computational requirements of these simulations make the project a perfect match for Olivia's powerful GPU nodes, enabling faster and more detailed analysis than ever before.
2. Computational Fluid Dynamics (CFD)
Alexander Hanke is using the open-source CFD model REEF3D for complex simulations, in the field of marine or coastal engineering. The ability to run larger models and get faster results on Olivia, compared to older systems like Betzy, allows for more detailed investigations and quicker research cycles.
3. Deep Learning for Geoscience
A project led by Benjamin Robson and Yiyao Chen at the University of Bergen aims to leverage deep learning for analysing vast, multi-terabyte geoscience datasets. Their work depends on a stable, high-performance environment that can handle large-scale data transfer and model training, capabilities for which Olivia is specifically designed.
4. AI for Agriculture
Nils Helset, CEO of DigiFarm, leads a project focused on training advanced machine learning models for agricultural applications. The significant speedup observed on Olivia enables the company to innovate more quickly, developing new tools to enhance farming efficiency and sustainability.