Maximum parallel throughput for training workloads

Best OCI Instances for ML Training

CPU-based machine learning training — gradient boosted trees (XGBoost, LightGBM), scikit-learn pipelines, and data preprocessing — scales linearly with parallel CPU throughput. The relevant metric here is peak PassMark at maximum vCPU count, not value per vCPU. Bare Metal shapes with 128–192 vCPUs let you parallelize cross-validation folds and hyperparameter search without network overhead. Flex shapes give you precise cost control: scale up for training runs, scale down for inference serving on the same shape family.