What are graph neural networks and why they matter

Most people know neural networks from text generation or image recognition. Graph neural networks (GNNs) are their specialized variant — they work with data that has a natural relational structure. Imagine a friendship map on Facebook, a network of banking transactions, or metabolic pathways in a cell. These structures cannot simply be "flattened" into a table or an image — you need a graph.

Today, GNNs power systems for fraud detection in banks, Netflix recommendation algorithms, weather forecasting, or drug discovery. The larger and more complex the graph, the more accurate the model — but also the greater the computational and memory demands on hardware.

The problem: GPUs can't handle massive graphs

Modern GPU cards are extraordinarily powerful for AI training — but they have one Achilles' heel: limited memory. While the top-tier NVIDIA H100 offers 80 GB of VRAM, real-world production graph datasets can have hundreds of millions of nodes and billions of edges. The entire graph simply won't fit into a GPU.

Previous solutions "addressed" this problem by moving some computations back to the processor (CPU). But that consumed most of the performance gain — the GPU waited, data was transferred over the slow bus, and training became expensive in both time and energy.

Capsule: smart memory management without compromises

The Data Darkness Lab team from the Medical Imaging Intelligence and Robotics Research Center at the University of Science and Technology of China (USTC) published a paper titled "Capsule: An Out-of-Core Training Mechanism for Colossal GNNs" in February 2025 in the journal Proceedings of the ACM on Management of Data.

The key idea: instead of "offloading" work to the CPU, Capsule intelligently divides the graph into smaller parts and ensures that each part fits precisely into the available GPU memory during processing. How does it do that?

• Graph partitioning and pruning — the system analyzes the graph structure and splits it into subgraphs to minimize redundant data.
• Loading order optimization — Capsule solves the so-called shortest Hamiltonian cycle (the traveling salesman problem applied to subgraphs) so that loading happens as efficiently as possible and data movement is minimized.
• Pipelining — while the GPU trains one subgraph, the CPU prepares the next one in the background. I/O operations and computation run in parallel.
• Elimination of CPU–GPU transfers during backpropagation — the critical training phase (backpropagation) stays entirely on the GPU, without costly data transfers.

Results: 12× faster, 4× less memory

Comparisons with the best existing systems speak clearly:

• Speed: up to 12.02× higher performance compared to state-of-the-art systems
• RAM: Capsule needs only 22.24% of the main memory consumed by competing systems
• Compatibility: the system works as a plug-and-play addition to the most widely used GNN frameworks — DGL (Deep Graph Library) and PyG (PyTorch Geometric)

Moreover, the researchers derived theoretical variance bounds for training embeddings — a mathematical guarantee that simplifying the graph does not reduce the quality of the resulting model.

The global race for efficient GNN training

Capsule isn't the only breakthrough in this field. A team from Korea's KAIST university published the FlexGNN system in summer 2025, achieving an even more dramatic speedup — up to 95× faster training on a single GPU card compared to existing technologies, by strategically moving data between GPU, RAM, and SSD in real time. It was published at the prestigious KDD 2025 conference.

Both projects — although using different technical approaches — target the same problem: democratizing the training of large GNN models without the need to invest in extremely expensive server clusters with dozens of GPU cards.

What this means for practice and Europe

Graph neural networks are increasingly in demand in industry. Banks deploy them for real-time fraud detection, healthcare facilities for analyzing patient data and genetic networks, logistics companies for route optimization. The problem has always been the same: the training infrastructure is too expensive for mid-sized companies.

If approaches like Capsule or FlexGNN make it into the mainstream (and both are integrable into existing PyTorch/DGL pipelines), it will open up GNN training even for smaller research teams and companies in Europe that don't have access to data centers full of expensive GPUs. For Czech and Slovak startups or research institutes, this could mean being able to train specialized GNN models — for example, for clinical data analysis or network anomaly detection — on significantly more modest hardware.

Capsule as such is an academic project — the code is not yet available as a finished product on a commercial marketplace. However, the researchers from USTC published the work openly through ACM and presented it at the SIGMOD 2025 conference.

What are DGL and PyG frameworks?

For those starting with GNNs: DGL (Deep Graph Library) and PyG (PyTorch Geometric) are the most widely used open-source libraries for working with graph neural networks. DGL is developed by Amazon Web Services, PyG originated at the Technical University of Munich and is now maintained by the community. Both libraries run on top of PyTorch and are completely free.