Data Science

NVIDIA CUDA-X Now Accelerates the Polars Data Processing Library

Decorative image.

Polars, one of the fastest-growing data analytics tools, has just crossed 9M monthly downloads. As a modern DataFrame library, it is designed for efficiently processing datasets that fit on a single machine, without the overhead and complexity of distributed computing systems that are required for massive-scale workloads.

As enterprises grapple with complex data problems—ranging from detecting time-boxed patterns in credit card transactions to managing quickly shifting inventory needs across a global customer base—even higher performance is essential.

Polars and NVIDIA engineers just released the Polars GPU engine powered by RAPIDS cuDF in open beta, bringing accelerated computing to the growing Polars community with zero code change required. This brings even more acceleration to the query execution for Polars, making this speedy data processing software up to 13x faster compared to running on CPUs. It’s like giving rocket fuel to a cheetah to help it sprint even faster.

“The collaboration with NVIDIA is a unique opportunity that offers the power of NVIDIA RAPIDS and GPUs to everyone looking to get even more performance from Polars,” said Ritchie Vink, author and CEO of Polars.

RAPIDS, part of NVIDIA CUDA-X, is an open-source suite of GPU-accelerated libraries designed to improve data science and analytics pipelines. RAPIDS cuDF is a GPU DataFrame library for loading, joining, aggregating, filtering, and manipulating data.

NVIDIA software accelerates data processing at every scale

With data science and engineering teams building more and more data processing pipelines to fuel AI applications, it’s critical to choose the right software and infrastructure for the job to keep things running smoothly.

For workloads well-suited to individual servers, workstations, and laptops, developers frequently use libraries like Polars to accelerate iterations, reduce complexity in development environments, and lower infrastructure costs. 

On these single-machine-sized workloads, quick iteration time is often the top priority, as data scientists often must do exploratory analysis to guide downstream model training or decision-making. Performance bottlenecks from CPU-only computing reduce productivity and can limit the number of test/train cycles that can be completed.  

For large-scale data processing workloads too large for a single machine, organizations turn to frameworks like Apache Spark to help them distribute the work across nodes in the data center. At this scale, cost and power efficiency are often the top priorities, but costs can quickly balloon due to the inefficiencies of using traditional CPU-based computing infrastructure.

The NVIDIA CUDA-X data processing platform is designed with these needs in mind, optimized for cost- and energy-efficiency for large-scale workloads and performance for single-machine-sized workloads.

Medium-scale workloads where productivity and performance are critical can see performance gains on both Polars, as well as 50x faster performance on the pandas library using NVIDIA GPU-enabled systems instead of CPUs, based on industry-standard benchmarks. 

With the RAPIDS Accelerator for Apache Spark, workflows where cost and energy efficiency are critical can see cost savings of up to 80% and energy savings of up to 12x.

Get started today

The world is creating more data than ever before and accelerated computing makes it possible to operationalize it efficiently. Whether you’re running on a workstation or scaling out in the data center, NVIDIA-accelerated data processing software can improve productivity and reduce costs.

For more information about how you can accelerate your data analytics workflows with zero code change, see the NVIDIA RAPIDS page.

Discuss (0)

Tags