IntelligentHQ
IntelligentHQ
Skip to content

Data Science & Analytics

GPU acceleration for dataframes, machine learning, graph analytics and optimization: CUDA-X Data Science (formerly RAPIDS) with cuDF, cuML and cuGraph, the RAPIDS Accelerator for Apache Spark, and the cuOpt decision optimization engine.

Technology profiles for this category are in research.

Overview

Data teams often spend more time waiting on joins, feature engineering and model training than on analysis. This category covers NVIDIA libraries that run familiar Python and Spark work on GPUs.

CUDA-X Data Science is the current name for the libraries long known as RAPIDS, and some parts, such as the RAPIDS Accelerator for Apache Spark, keep the old name. cuDF speeds up dataframe work and offers zero-code-change acceleration for pandas, Polars and Spark. cuML accelerates scikit-learn style algorithms, UMAP and HDBSCAN. cuGraph accelerates NetworkX graph analytics. cuOpt is an open-source engine for vehicle routing, linear programming and mixed-integer programming, with MILP support in beta.

The audience is data scientists and data engineers whose jobs take minutes or hours on CPUs.

For small tables that already finish in seconds, a GPU adds setup cost for little gain. Check that your operations are supported before migrating a pipeline.12

Problems it addresses

  • Slow dataframe pipelines1

    Joins and group-bys on large tables take long on CPUs. cuDF runs pandas and Polars code on GPUs without rewrites.

  • Long training on tabular data1

    cuML accelerates scikit-learn style algorithms, UMAP and HDBSCAN.

  • Costly Spark clusters1

    The RAPIDS Accelerator for Apache Spark moves Spark data processing onto GPUs.

  • Graph analysis at scale1

    cuGraph accelerates NetworkX workloads.

  • Routing and planning decisions2

    cuOpt targets routing and linear programs with millions of variables and constraints.

A typical workflow

  1. Profile the CPU pipeline

    Time each stage and find the slowest operations.

  2. Turn on GPU acceleration1

    Load the cuDF pandas accelerator or Polars GPU engine and rerun the unchanged code.

  3. Train models1

    Move scikit-learn style training to cuML.

  4. Scale out1

    Use Dask or the RAPIDS Accelerator for Apache Spark for multi-node jobs.

  5. Optimize decisions2

    Feed results into cuOpt for routing or planning problems.

NVIDIA Solution Architect

Describe your project and get an explainable architecture.

Open the lab

Next steps

  1. Run your slowest notebook with the cuDF pandas accelerator on one GPU and compare wall-clock time.

  2. Check the supported-operations list for your libraries before planning a migration.

  3. For routing or scheduling, install cuOpt from PIP, Conda or a container and test it on a sample instance.

  4. Use the NVIDIA Solution Architect if analytics is one part of a wider AI project.

Sources

  1. NVIDIA CUDA-X for Data Science (opens in a new tab)NVIDIA · Vendor-reported
  2. NVIDIA cuOpt (product page) (opens in a new tab)NVIDIA · Vendor-reported

Fill out the form below to request your copy.

Name(Required)