At LifeSpan AI, two high-performance servers provide the foundation for developing complex AI models and thoroughly testing new methods. It’s not just about speed; it’s also about determining which research questions can be addressed in the first place.
How does health change over the course of a lifetime? What signs foreshadow diseases? And how can insights be gained from data to enable more targeted prevention? The DFG research group LifeSpan AI develops artificial intelligence methods for these purposes. The research team combines information from various sources, examines individual health trajectories, and looks for ways to distinguish statistical correlations from causal relationships.
When data contains many features or comes from different sources, observations span long periods of time, and models are expected to learn complex relationships, the demands on computing power and memory increase. This is also the case in health research. LifeSpan AI therefore relies on two complementary systems: a GPU server for deep learning and a CPU server for extensive parallel computations.
Recognizing Patterns: The GPU Server
The GPU server features eight NVIDIA RTX 6000 Ada Generation graphics processing units, each with 48 gigabytes of dedicated memory. They handle the many repetitive computational operations involved in training and applying neural networks. This is particularly important for embedding models being developed for different data modalities over the course of the project. These models translate, for example, tabular data, sensor data, or images into numerical representations that AI models can use for further processing. The aim is to map relevant structures—such as similarities between observations that are not immediately apparent in the original data—and to provide a basis for making predictions for new patients as accurately as possible.
The large amount of memory per GPU provides space for sophisticated models and the data they need during their computations. Multiple GPUs also make it possible to distribute computational tasks depending on the method or to run different training runs simultaneously. Two AMD EPYC processors and 1.5 terabytes of memory support the processing and provision of the data.
Putting Methods to the Test: The CPU Server
Developing an AI model is only part of the work. Equally important is the question: Under what conditions does a method deliver reliable results—and where are its limitations? Simulation studies are an important tool for this. Researchers generate artificial data whose properties they know and use it to test their methods. This allows them, for example, to examine how a method reacts to missing measurements, samples of varying sizes, or complex relationships. Such controlled tests are particularly valuable when methods are intended not only to make predictions but also to provide insights into causes.
This is where the CPU server comes into its own. Its two processors have a combined total of 128 cores and support up to 256 concurrently processed execution threads. This allows many independent simulation runs to be processed in parallel. This server also features 1.5 terabytes of RAM—enough capacity to keep extensive data and interim results in memory.
Developing and Validating—on a Solid Foundation
The two servers perform different tasks but serve the same goal: to develop new AI methods for health research and systematically evaluate their validity. The GPU server supports the learning of complex patterns. The CPU server enables broad methodological testing involving many independent calculations.
Computing power alone does not make a result reliable. However, it creates the conditions for comparing more variants, testing assumptions more thoroughly, and addressing complex research questions in practice. In this way, the infrastructure supports LifeSpan AI’s aim to better understand health across the lifespan and lay the groundwork for more targeted prevention.
Technical Specifications at a Glance
GPU Server:
CPU Server: