AMD and Intel merge libraries to streamline AI inference on x86 CPUs through unified PyTorch path

AMD and Intel Merge CPU Optimization Libraries for Unified x86 AI Development

AMD and Intel have announced a significant collaboration to integrate their CPU optimizations for PyTorch into a single library, marking a strategic shift in how AI inference on CPUs will be handled across x86 architectures1

1

2

2

. The AMD ZenDNN library, previously maintained as a downstream version of Intel's oneDNN, is now being folded into the open-source oneDNN framework published by the UXL Foundation. This move creates a unified PyTorch code path that provides optimized tuning for both AMD and Intel chips, eliminating the need for developers to maintain separate software paths for different x86 CPUs.

x86 Ecosystem Advisory Group Drives Collaboration

The integration was recommended by the x86 Ecosystem Advisory Group (EAG), a joint initiative established in 2024 primarily between AMD and Intel to advance the x86 ecosystem1

1

. The two rivals have previously collaborated through the EAG on initiatives like the ACE extensions for x86. AMD has now joined the UXL Foundation as a steering member, contributing $20,000 per year to influence the foundation's future direction1

1

. This membership solidifies AMD's commitment to advancing open-source library development for x86 acceleration alongside founding members Intel, Arm, Qualcomm, and Google Cloud.

Technical Performance and Developer Benefits

The oneDNN backend is specifically designed for CPU inference and accelerates AI workloads on the latest Intel and AMD chips within frameworks like PyTorch1

1

. When oneDNN was added to PyTorch as a backend in late 2023, it demonstrated a 1.24x improvement with the Float32 data type and 2.21x performance boost for the BFloat16 data type1

1

. The library detects hardware features at runtime to generate optimized kernels for the platform and provides low-precision data types along with optimized instructions for common AI operations such as matrix multiplication and convolution.

Jeff McVeigh, Intel's vice president and general manager of Data Center Strategic Programs, emphasized the practical impact: "Developers shouldn't have to think about separate software paths to get the best from an x86 CPU. Working with AMD through oneDNN gives the PyTorch community a consistent approach for modern processors"2

2

.

Reducing Ecosystem Fragmentation

This collaboration addresses a long-standing challenge in the x86 ecosystem by simplifying AI development across different CPU vendors2

2

. Robert Hormuth, AMD's corporate vice president of Architecture and Strategy, stated that AMD is "advancing an open software ecosystem that gives developers greater choice and makes it easier to build high-performance AI applications"2

2

. For developers working with PyTorch, the unified approach ensures a single backend provides proper optimizations regardless of the underlying hardware, reducing time spent on vendor-specific workarounds and allowing more focus on application development.

Cross-Platform Reach and Future Implications

While the announcement focuses on x86 CPUs, oneDNN functions as a cross-platform library that also works on Arm and RISC-V architectures1

1

. This broader compatibility positions the library as a versatile solution for AI workloads across multiple processor architectures. The collaboration strengthens CPU performance for AI inference at a time when organizations are scaling their AI deployments and seeking greater flexibility in hardware choices. By aligning around a common foundation, AMD and Intel are removing friction points that have historically complicated development across x86 platforms, potentially accelerating adoption of CPU-based AI inference workloads.