Rongchai Wang Aug 25, 2026 18:43

Anyscale's Ray Data now supports NVIDIA cuDF and RapidsMPF, delivering up to 3x better TCO for GPU-accelerated AI data processing workloads.

Anyscale has announced a major update to Ray Data, its scalable data-processing library for AI workloads, by integrating GPU-native operators in collaboration with NVIDIA. Leveraging NVIDIA's cuDF and RapidsMPF, Ray Data now promises up to 3x better total cost of ownership (TCO) and enhanced performance for GPU-accelerated data curation tasks, according to the company’s latest blog post.

Ray Data, part of the broader Ray ecosystem, is designed to handle the increasingly demanding data-processing needs of AI projects, such as multimodal data handling and training data pipelines. Prior to this update, Ray Data supported GPU scheduling but lacked native functionality for executing directly on GPUs. The new integration addresses this limitation, enabling faster and more cost-effective processing for compute-intensive workloads.

Key Updates: cuDF and RapidsMPF

The integration introduces two critical features:

• cuDF Batch Format Support: Ray Data users can now process data using NVIDIA’s cuDF, a GPU-accelerated DataFrame library with a Pandas-like API. This eliminates the need for manual format conversions and speeds up operations by several orders of magnitude compared to CPU alternatives. In benchmarking tests, GPU-accelerated data hashing via cuDF was up to 15x faster than a CPU baseline on the FineWeb 10BT dataset.
• GPU-Native Shuffle with RapidsMPF: RapidsMPF facilitates high-speed data shuffling by bypassing CPU memory. This is particularly impactful for tasks like group-by operations, often bottlenecks in AI data pipelines. For shuffle-intensive workloads, the GPU solution demonstrated speedups of 3x over CPU-based approaches, though diminishing returns were noted at higher GPU scales due to communication overheads.

Implications for AI Workloads

The updates target AI and machine-learning pipelines that require massive data preprocessing ahead of model training. For instance, a fuzzy document deduplication benchmark on Ray Data showed GPU acceleration reduced end-to-end pipeline times by 4x while lowering TCO by 3.1x on a two-node cluster. These gains are attributed to the GPU's ability to handle computationally heavy tasks like MinHash signature generation and iterative connected components detection more efficiently than CPUs.

These improvements align with broader trends in AI infrastructure, where hybrid cloud and accelerator adoption are driving demand for optimized processing libraries. Anyscale's recent partnership with Nscale, announced in July 2026, further signals its commitment to optimizing Ray for next-generation data-center architectures.

What’s Next for Ray Data

While the current updates mark a substantial step forward, Anyscale hinted at further enhancements. Plans include operator fusion to reduce host-to-device memory transfers, which would unlock even better performance for sequential tasks. The company is also working to extend GPU-native shuffling to Ray Data’s preprocessor API, enabling more feature-engineering pipelines to benefit from GPU acceleration.

Ray Data’s GPU-native operators are now available in the experimental release of Ray 2.58, which can be installed via pip install -U ray==2.58. As AI workloads continue to grow in size and complexity, these advancements position Ray Data as a key enabler for scalable, cost-efficient data pipelines.

Image source: Shutterstock

read more