# Databricks releases beta loader that streams governed tables to training GPUs

_Published Monday, November 9, 2026 at 10:43 AM EST · AI, Infrastructure · Latest · Tier 2 — Notable_

![Databricks releases beta loader that streams governed tables to training GPUs — Primary](https://www.databricks.com/sites/default/files/blog_images/accelerate-gpu-training-on-unity-catalog-tables-blog-img-og.png)

Databricks has released UCDataset in beta on AI Runtime, allowing PyTorch training jobs to stream Delta tables from Unity Catalog directly to GPUs without exporting a separate copy. The loader divides matching rows across GPUs and data-loader workers, removing the need for custom code to allocate those rows.

Databricks reports a 10x improvement in training time against loading Parquet data in a simulated production workload using a 300 GB feature table and 8 NVIDIA H100 GPUs. Jobs use the running identity's table permissions. The beta cannot read tables with row filters or column masks; support for images, audio and documents through table files is still being expanded.

## Sources

- [Databricks](https://www.databricks.com/blog/accelerate-gpu-training-unity-catalog-tables-ucdataset)

---
Canonical: https://techandbusiness.org/newswire/BSYK_iOrNxuXIlGEvlAv45
Published: 2026-11-09T15:43:56.418Z
Story chronology: 2026-11-09T15:00:00.000Z
Retrieved: 2026-11-09T17:47:08.883Z
Publisher: Tech & Business (techandbusiness.org)