Intel oneDNN AI Drives TensorFlow 2.9

May 27, 2022

Intel and Google team up to enable the oneDNN library as the default backend CPU optimization for TensorFlow 2.9.Workflow Orchestration Demand Grows

In the latest release of TensorFlow 2.9, the performance improvements delivered by the Intel oneAPI Deep Neural Network Library (oneDNN) are turned on by default. This applies to all Linux x86 packages and for CPUs with neural-network-focused hardware features (like AVX512_VNNI, AVX512_BF16, and AMX vector and matrix extensions that maximize AI performance through efficient compute resource usage, improved cache utilization and efficient numeric formatting) found on 2nd Gen Intel Xeon Scalable processors and newer CPUs. These optimizations enabled by oneDNN accelerate key performance-intensive operations such as convolution, matrix multiplication and batch normalization, with up to 3 times performance improvements compared to versions without oneDNN acceleration.

Wei Li, Intel vice president and general manager of AI and Analytics said, “Thanks to the years of close engineering collaboration between Intel and Google, optimizations in the oneDNN library are now default for x86 CPU packages in TensorFlow. This brings significant performance acceleration to the work of millions of TensorFlow developers without the need for them to change any of their code. This is a critical step to deliver faster AI inference and training and will help drive AI Everywhere.”

oneDNN performance improvements becoming available by default in the official TensorFlow 2.9 release will enable millions of developers who already use TensorFlow to seamlessly benefit from Intel software acceleration, leading to productivity gains, faster time to train and efficient utilization of compute. Additional TensorFlow-based applications, including TensorFlow Extended, TensorFlow Hub and TensorFlow Serving also have the oneDNN optimizations. TensorFlow has included experimental support for oneDNN since TensorFlow 2.5.

oneDNN is an open source cross-platform performance library of basic deep learning building blocks intended for developers of deep learning applications and frameworks. The applications and frameworks that are enabled by it can then be used by deep learning practitioners. oneDNN is part of oneAPI, an open, standards-based, unified programming model for use across CPUs as well as GPUs and other AI accelerators.

While there is an emphasis placed on AI accelerators like GPUs for machine learning and, in particular, deep learning, CPUs continue to play a large role across all stages of the AI workflow. Intel’s extensive software-enabling work makes AI frameworks, such as the TensorFlow platform, and a wide range of AI applications run faster on Intel hardware that is ubiquitous across most personal devices, workstations and data centers. Intel’s rich portfolio of optimized libraries, frameworks and tools serves end-to-end AI development and deployment needs while being built on the foundation of oneAPI.

The oneDNN-driven accelerations to TensorFlow deliver remarkable performance gains that benefit applications spanning natural language processing, image and object recognition, autonomous vehicles, fraud detection, medical diagnosis and treatment and others.

Deep learning and machine learning applications have exploded in number due to increases in processing power, data availability and advanced algorithms. TensorFlow has been one of the world’s most popular platforms for AI application development with over 100 million downloads. Intel-optimized TensorFlow is available both as a standalone component and through the Intel® oneAPI AI Analytics Toolkit, and is already being used across a broad range of industry applications including the Google Health project, animation filmmaking at Laika Studios, language translation at Lilt, natural language processing at IBM Watson and many others.