NVIDIA XGBoost 3.0: Training Terabyte-Scale Datasets with Grace Hopper Superchip

NVIDIA has released XGBoost 3.0, enabling training of gradient-boosted decision tree models on datasets up to 1 terabyte using a single GH200 Grace Hopper Superchip. This breakthrough leverages the new External-Memory Quantile DMatrix and the chips coherent memory architecture with 900GB/s NVLink-C2C bandwidth to stream compressed data directly from host RAM to GPU, overcoming previous memory limitations and simplifying large-scale machine learning workflows.