FedML's support for NVIDIA GPUs

FedML (fedml-dsp) can be used in notebooks with GPU compute. FedML now has support for NVIDIA RAPIDS™ and CUDA cuDF and cuML and detection of GPU for adding support for RAPIDS™ CUDA.

FedML's Connectivity Core component supports reading data from semantic models of SAP Datasphere directly into CUDA cuDF dataframes. FedML's GPU support helps data scientists accelerate the machine learning workflows with NVIDIA GPUs, while providing instant access to SAP's critical business data without the need for ETL or additional overhead in processing ETL'd data.

Architecture

Solution Diagram Resources

You can download the Solution Diagram as a .drawio file for offline use. Alternatively, you may view and edit the Solution Diagram directly on draw.io.
Please note that any changes made online will need to be saved locally if you wish to keep them.

Download

Edit

Resources

Pip installable library: https://pypi.org/project/fedml-dsp/

When and Where to use

When models need to be trained (and feature calculations performed) using complex and intensive algorithms (using SAP business data), they require the processing power of NVIDIA GPUs for efficient model training. Some performance acceleration numbers compared between CPU and GPU training noted below.
When critical SAP data from various SAP applications (with semantics intact) is needed for such GPU model training jobs, FedML helps read the SAP data (via SAP Datasphere semantic models) directly into RAPIDS™ supported cuDF format, in real-time, without the need to copy the data to the machine learning platform via ETL first.
Models trained with NVIDIA GPUs can also be deployed in SAP AI Core (GPU deployments) for quick inferencing with SAP business data.

CPU-GPU Performance Acceleration metrics

Comparing several cuML algorithms on GPUs vs their CPU equivalents, the following acceleration numbers were recorded for GPU environments with varying data loads and features:

RandomForest: