XGBoost GPU Acceleration Updates Boost Speed and Multi‑GPU Scalability

The XGBoost library has received major GPU‑based enhancements. A histogram‑based decision‑tree construction method replaces the previous linear‑time search, delivering significantly faster training while preserving accuracy. Multi‑GPU support now leverages the NCCL library, enabling scalable communication across devices and allowing workloads such as the 115 million‑row airline dataset to run efficiently on AWS P3 instances with up to eight GPUs.

Earlier releases introduced a GPU plugin that provided large speedups over multicore CPUs for large structured‑data tasks, correcting benchmark errors and confirming the benefits for the machine‑learning community. These updates continue to improve memory efficiency and runtime for gradient‑boosting models on modern hardware.