Databricks takes AI Runtime GA with expanded NVIDIA tooling

Databricks has made AI Runtime generally available, adding NVIDIA NeMo, NGC containers and new Nemotron cookbooks for training and fine-tuning open models.

Developers can use AI Runtime for multi-node distributed training on NVIDIA GPUs while keeping workloads alongside enterprise data. Databricks says it handles the underlying clusters, networking and reliability, reducing the infrastructure work required of model teams. The company did not provide an independent benchmark quantifying that reduction.

The release follows AI Runtime’s public-preview launch on March 19. During the preview, Databricks offered single-node distributed training across eight H100 GPUs, while multi-node support remained in private preview. Databricks says the generally available system uses NVIDIA Quantum InfiniBand for remote direct memory access networking and NCCL for communication between GPUs.

The NVIDIA integrations span several parts of the model-development stack. According to NVIDIA’s documentation, NeMo Framework provides tools for building and customizing generative AI models, including parallel and distributed training methods. Databricks says AI Runtime can run NeMo and NeMo AutoModel on its serverless GPUs. The Nemotron developer asset hub includes training recipes, usage cookbooks, datasets and end-to-end examples for NVIDIA’s open-model family.

AI Runtime also supports Docker and private registries. Developers can bring GPU-optimized NGC containers into the service through Databricks Artifact Registry. Databricks says environment caching is intended to reduce repeated setup time. NVIDIA describes NGC as a catalog of containers, pretrained models and other GPU-optimized resources; some catalog items require sign-in, entitlements or acceptance of license terms.

Databricks says its AI Runtime product uses Unity Catalog to govern access to training data, MLflow to track experiments and GPU utilization, and system tables to expose CPU and GPU usage. Databricks and NVIDIA say that combining those controls with NVIDIA tooling and accelerated compute can help teams use enterprise data to create specialized models and agents. The reviewed sources do not include an independent customer evaluation showing faster customization or better model outcomes.

The AI Runtime quickstart explains how users can submit and monitor serverless GPU workloads through the Databricks command-line interface, with example configurations requesting one A10 GPU or one H100 GPU.