ST PAUL, Minn., Nov. 21, 2024 — Hyperion Research today announced the recipients of the 20th round of HPC Innovation Excellence Awards.
Since 2011, these semi-annual innovation excellence awards have been recognized globally as the touchstone for outstanding achievements that were supported by the use of high performance computing. Winners have ranged from government and academic research teams to some of the world’s largest corporations. The program’s goals are to showcase HPC-supported, real-world achievements with significant potential for benefiting humanity — achievements that demonstrate the value of HPC for research and development in government, academic or private-sector organizations.
Only a few of the many nominations receive these awards. Judges for the awards, chartered with carefully reviewing each submitted nomination, are the members of the worldwide steering committee of the HPC User Forum, representing leading HPC user organizations in government, academia and industry. Award winners receive a $1,000 honorarium and an award certificate and trophy recognizing their achievements.
The 2024 winners of the HPC Innovation Awards are briefly described here:
AxoNN: Democratizing AI via Open-source Scalable AI Training and Fine-tuning
- Organization: University of Maryland
- Contact: Siddharth Singh and Abhinav Bhatele
- Innovation: AxoNN is a scalable, highly parallel AI framework for training and fine-tuning that uses a novel 4D hybrid parallel algorithm. Selected as a Gordon Bell award finalist for SC ’24, AxoNN leverages Agarwal’s 3D matrix multiplication approach. Large Language Model (LLM) training using AxoNN achieves over 620 Petaflop/s on NVIDIA A100 GPUs, 1423 Petaflop/s on H100 GPUs, and 1381 Petaflop/s on AMD MI250X GPUs—outperforming current benchmarks by over 25%. However, as LLMs grow, so do risks of privacy breaches due to data memorization. One such study using AxoNN highlights these concerns by exploring “catastrophic memorization,” where models can memorize data in a single pass, and proposes mitigation strategies. The framework has also demonstrated fine-tuning a 405-billion parameter LLM on the Frontier supercomputer, addressing both performance and privacy challenges.
Closed-Loop Mars Entry Simulations with Distributed Exascale Computing
- Organization: NASA
- Contact: Eric Nielsen
- Innovation: Human-scale Mars missions will require large landers using retropropulsion, as traditional parachutes are infeasible for the massive payloads needed. With no prior experience and limited testing possible on Earth, high-fidelity HPC simulations will be critical for developing and validating new technologies. Researchers from NASA and Georgia Tech executed unprecedented, autonomous flight trajectory simulations using the Summit and Frontier platforms at Oak Ridge National Laboratory, demonstrating a viable approach for simulating future crewed Mars missions. The software developed has broad applications across aerospace, defense, and academia. This work has received recognition in multiple forums and media coverage.
Advancing precision and resilient agriculture with AI, reducing costs and boosting crop yields
- Organization: DigiFarm
- Contact: Nils Helset
- Innovation: DigiFarm uses AI models, trained on LUMI, to accurately detect field boundaries and seeded acres, surpassing human accuracy. By super-resolving Sentinel-2 satellite imagery to 1-meter resolution and utilizing 4.7 million hectares of training data from 57 countries, DigiFarm achieves detection accuracies above 94%, which is 12-15% better than existing methods. Its commercial SaaS/API solution, launched in 2022, now serves over 50 corporate clients, including Bayer, Yara, CNHi, ENI, Syngenta, and various governments. The solution reduces agricultural monitoring costs by over 25%, saving Lithuania, for instance, €1.5 million annually.
Hyperion Research welcomes award entries from around the world. Entries may be submitted at any time by completing the application form available at https://www.hpcuserforum.com/innovationaward/. Submissions must contain a clear description of the impact of the project. The HPC User Forum Steering Committee performs an initial ranking of submissions, after which domain and vertical experts are called on, as needed, to evaluate the submissions.
