Anaconda Inc. announced today it is leveraging Docker containers and Kubernetes clusters to accelerate the development of AI applications built and deployed using graphical processor units (GPUs) from NVIDIA.
Previously, Anaconda added support for Docker and Kubernetes to version 5.0 of Anaconda Enterprise, a commercially supported instance of an open source platform for developing, governing and automating data science and AI pipelines on Intel processors. A version 5.2 of Anaconda Enterprise extends that platform to add support for GPUs.
Matthew Lodge, senior vice president of products and marketing at Anaconda, says that training AI applications has been proven to be significantly faster on GPUs. But over time, developers of AI applications will be employing a broad range of algorithms across Intel processors, GPUs, field programmable gate arrays and new classes of processors such as the TPU processors developed by Google, which are designed specifically for AI applications. Anaconda provides a mechanism for managing the building of multiple AI models that eventually will span all those processor classes thanks to the portability enabled by Docker and Kubernetes, Lodge says.
The Anaconda platform is currently being employed by more than 6 million data scientists. Lodge says the platform is designed to make it simple to develop score models on a subset of data that then can be scaled up to support model development to a 1,000-node GPU cluster for training, and then deployment on a production deployment. Lodge says data scientists and developers favor Anaconda because it allows them to model an AI application using a laptop before having to commit to using a cloud service.
Lodge says it’s still early days in terms of crafting best DevOps practices for building and deploying AI applications. The biggest challenge with building an AI application is not necessarily the applications as much as it is training the algorithms to identify and correctly respond to patterns. The more data an algorithm is exposed to, the more accurate it becomes, which is why most AI applications today are deployed on GPUs that make it possible to access larger amounts of data faster.
It may be a while before data scientists and DevOps teams come together to create consistent processes for developing and training AI models. That said, most organizations still don’t appreciate the fact that AI models will need to be updated and replaced as the number of data sources that algorithms can be applied against continues to expand. Without some structured means of managing those AI models, organizations could apply algorithms that no longer are relevant to automate a process.
In the meantime, DevOps teams should start to engage data scientists now. It’s only now a matter of time before every enterprise application is augmented using AI models that need to be maintained over the life cycle of application. In some ways managing those AI models will be the same as it ever was to DevOps teams. The challenge now is teaching the data scientists that build these applications how to work as part of a larger team before any bad habits take root.