Technology – An Overview Accelerating Machine Learning Workloads with Open XLA

Google’s Open XLA project is an open-source software library for accelerating linear algebra computations on CPUs and GPUs. It started in 2017 as a collaboration between Google and TensorFlow teams, with the aim of providing a high-performance linear algebra library for machine learning workloads.

Initially sponsored by Google, and supported by contributions from AI/ML leaders including AMD, Arm, Intel, Meta, and NVIDIA, Open XLA has grown to be a community-driven project with contributions from developers around the world. It is built upon several other open-source projects, including Eigen, BLAS, and cuBLAS.

The library supports a wide range of CPUs and GPUs, including Intel CPUs, NVIDIA GPUs, and AMD GPUs. In addition to supporting these hardware platforms, Open XLA provides a unified programming interface that abstracts away hardware-specific details, making it easy for developers to write high-performance machine learning code.

Open XLA is also integrated with several other popular machine learning frameworks, including TensorFlow, PyTorch, and JAX. This integration allows developers to use Open XLA with their existing machine learning code and easily incorporate it into their workflows.

One of the key benefits of the Open XLA approach is that it enables the efficient use of hardware resources, resulting in faster training and inference times for machine learning models. For example, a study conducted by Google showed that using Open XLA to train a ResNet-50 model on a TPU v3 resulted in a 1.8x speedup compared to using TensorFlow without Open XLA.

The growth in contributions to the project has been impressive. As of March 2023, there have been 1,184 commits from 70 contributors, with over 9,000 lines of code added and over 3,000 lines of code removed. The project has a vibrant community, with active discussion and support channels on GitHub and Slack.