Intel and AMD unify AI tools to ease x86 model use

Intel and AMD announced a collaboration to integrate Intel's oneDNN with AMD's ZenDNN into a more unified programming interface to simplify running AI models on x86 processors.

Intel confirmed it will coordinate its oneDNN libraries with AMD's ZenDNN. The companies said the aim is to simplify PyTorch support so developers need fewer separate software paths to support processors from both vendors.

oneDNN is a set of optimized libraries for deep learning operations on processors and other hardware. ZenDNN is AMD's set of software tools for optimizing deep learning on AMD CPUs.

The firms said harmonizing these tools should reduce software fragmentation and make maintenance and development of AI applications easier.

Beyond software, Intel and AMD are working on ACE, or Advanced Matrix Extensions, targeted at improving matrix computations used in language models and neural networks on x86 processors.

Hardware support for ACE is not yet available. The companies expect ACE to appear in hardware later this decade, so practical benefits will depend on the arrival of compatible processors and supporting software.

The collaboration could matter for enterprise and low-latency systems where running models on CPUs rather than expensive accelerators may be preferable when latency, resource consumption and hardware cost are key concerns.

Despite long-standing competition between Intel and AMD, joint work on standards and tools could benefit developers. The ultimate effect on AI model performance will depend on final implementation details and independent test results.