Databricks has made AI Runtime generally available, giving customers access to managed GPUs for model training within its data platform. The company says teams can run training jobs without provisioning and maintaining a separate GPU system.
AI Runtime offers serverless, on-demand GPU access for distributed training. Databricks says it handles GPU procurement, cluster configuration, networking and health checks. It says workloads can scale to hundreds of GPUs, and that even clusters with dozens of GPUs can start in seconds.
Teams can use existing training tools
Developers can submit workloads through a command-line interface or work in notebooks and IDEs. They can use prepared environments or their own containers, including for Ray workloads. Databricks says environment caching avoids rebuilding the same setup for each experiment.
The release also adds an MLflow view for logs from training across multiple machines. Databricks says logs begin streaming seconds after submission. AI Runtime integrates with Lakeflow for coordinating training with data pipelines, while billing system tables show usage and costs by workload.
Unity Catalog governs GPU workloads
According to Databricks, Unity Catalog applies read and write permissions to training data, checkpoints and logs. The company says AI Runtime also provides virtual machine isolation, PrivateLink support and controls over outbound network access.
Databricks says hundreds of companies already use AI Runtime, including Robinhood, PicPay, Merck, Flo Health, TinyFish and FactSet. Superhuman reported moving a training workload from one machine with eight A100 GPUs to multiple machines with 16 H100 GPUs, cutting its run from roughly seven days to 48 hours.
AI Runtime is generally available now. Databricks provides a quickstart, template notebooks and starter guides for customers beginning a GPU workload.
