KI-Cluster-Upgrade-DGX v2

The subject of the contract is the delivery of an NVIDIA DGX B300 system. For the dtec.bw MissionLab project, UniBw M already carried out a procurement procedure in 2023 for the expansion of the existing deep learning infrastructure. In order to continue using the installed and established software, including the training software framework, training data, and already pre-trained neural networks, and to ensure common and efficient use of the previous deep learning workstation, further DGX-H100 systems and the corresponding NVIDIA network infrastructure were necessary. UniBw M awarded the contract on 16.12.2023 after conducting a negotiated procedure without a prior call for competition ("Delivery, installation and networking of a DeepLearning cluster (NU 187)") by awarding to DELTA Computer Products GmbH (Delta). Based on orders NU187, MT990, LT218, UniBw M received a Deep-Learning-High-Performance-GPU cluster within the scope of the MissionLab project (dtec.bw). Delta installed and networked five DGX systems on the UniBw M campus and integrated a high-performance/high-storage server. UniBw M now operates a powerful system with 1,120 logical processing cores (CPUs), ten terabytes of RAM, 40 NVIDIA H100 Tensor Core GPUs, networking of the five servers via Infiniband, and a high-performance/high-storage data server connected via NVLink. The 40 NVIDIA H100 Tensor Core GPUs (DGX systems) are also connected via NVLink. UniBw M uses its GPU cluster for, among other things, training Large Language Models, mission-related planning using Reinforcement Learning, and for training sensor deployment models and automated sensor evaluation (Neural Networks, Vision Transformers). The research work in the "MissionLab" project generates a constantly growing demand for powerful computing infrastructure. Automated mission systems must be simulated, evaluated, and further developed in realistic scenarios. These investigations are based on large amounts of data, complex AI models, and iterative training processes that conventional server structures can no longer reliably map. Long computation times, limited graphics memory, and lack of scalability would significantly slow down the execution of demanding simulations and the training of modern AI models, thus jeopardizing the project's objectives. Without the required update, not only the research goals of MissionLab but also other dtec projects would be endangered. To close this gap, UniBw M must procure a system that can permanently handle high computing and storage loads, provides a stable, certified AI platform, and can be seamlessly integrated into the existing research infrastructure.