FloatSOM: GPU-Accelerated, Distributed, Topology-Flexible Self-Organizing Maps
A new GPU-accelerated SOM framework handles datasets too large for device memory...
FloatSOM, developed by Tony Xu, Sarah Klamt, Katherine Turner, Anne Brustle, Felix Marsh-Wakefield, and Givanna Putri, tackles a critical scaling bottleneck in Self-Organizing Map (SOM) analysis: datasets that exceed GPU device memory. The framework introduces multi-GPU execution, out-of-memory disk-backed streaming, and novel topologies beyond standard rectangular or hexagonal lattices. This allows users to train SOMs on massive datasets without requiring expensive hardware upgrades, as data can be streamed from disk during training.
In rigorous evaluation across 14 synthetic and real benchmark datasets, FloatSOM demonstrated superior performance. Its topology-aware hyperparameter fine-tuning reduced quantization error compared to current state-of-the-art SOM implementations. The most impressive result: a 1024-node SOM network trained on 1,000,000,000 samples with 50 features completed in just 6.16 minutes across 8 GPUs spanning two separate high-performance computing nodes. This makes FloatSOM a practical solution for large-scale clustering and visualization tasks in fields like genomics, image analysis, and sensor data processing.
- FloatSOM supports multi-GPU execution and out-of-memory disk-backed streaming for datasets exceeding GPU memory limits.
- Novel topologies beyond regular lattices, combined with topology-aware hyperparameter tuning, reduce quantization error vs. state-of-the-art baselines.
- Trained a 1024-node SOM on 1 billion samples with 50 features in 6.16 minutes using 8 GPUs across two HPC nodes.
Why It Matters
FloatSOM enables scalable, high-throughput SOM training on massive datasets previously limited by GPU memory constraints.