At the GPU Technology Conference in Munich, Germany, Nvidia, a leader in high-performance GPUs and artificial intelligence made another step with the announcement of a new set of open source RAPIDS libraries for GPU-accelerated analysis and machine learning.
Nvidia RAPIDS, Open Source Libraries for IA
This time Nvidia is not announcing a new GPU platform or a new proprietary SDK for in-depth learning, but a new set of open source libraries for GPU-accelerated analysis and machine learning. The new libraries nicknamed RAPIDS will provide Python interfaces similar to those of Scikit Learn and Pandas, but will leverage the company’s CUDA platform to accelerate on one or more GPUs.
According to Jensen Huang, CEO of Nvidia, who on Tuesday informed several technology journalists by phone, Nvidia has experienced 50 times shorter training time using RAPIDS instead of a pure CPU implementation. This speed was measured in scenarios with the XGBoost ML algorithm in a Nvidia DGX-2 system, although the hardware configuration of the CPU was not explicitly discussed.
RAPIDS apparently includes Apache Arrow memory column data technology and is designed to run on Apache Spark. In this context, the company has purchased software from Databricks that will integrate RAPIDS into its own analysis and IA platform.
However, Databricks is not the only big name behind the RAPIDS platform. Technology giants such as IBM, Hewlett Packard Enterprise and Oracle are also in action.