AMD ZenDNN 6.1 adds LowOHA inference kernels for INT8, BF16, and FP16

AMD has released ZenDNN 6.1, an update to its oneDNN API-compatible deep learning library optimized for Zen processors. The release arrives one month after AMD and Intel announced joint collaboration on PyTorch and oneDNN optimizations, which produced oneDNN 3.14 with initial Zen support through the UXL Foundation. Despite that partnership, AMD is maintaining independent development of ZenDNN with this release.

ZenDNN 6.1 focuses on advancing the LowOHA inference stack with new custom kernels. These include faster group MATMUL operations, FP16 and DA8W4 kernels, and additional MATMUL kernels for DA8W4, DA8W8, INT8, and BF16 data types. The release also introduces symmetric INT8 prepacking and other inference-oriented improvements. ZenDNN 6.0 shipped in July 2026 with broader enhancements, making this a targeted follow-up for inference performance.

What changes in ZenDNN 6.1

• LowOHA inference stack: The release extends AMD's inference-optimized stack with new kernel variants designed for reduced-precision matrix operations on Zen architectures.
• Group MATMUL kernels: Faster custom kernels for grouped matrix multiplication are included, a common pattern in transformer and convolutional inference workloads.
• Quantized data type support: New kernels cover DA8W4 (8-bit activation, 4-bit weight), DA8W8, INT8, and BF16 formats, enabling compressed model execution with reduced memory bandwidth.
• Symmetric INT8 prepacking: Added optimization for INT8 inference that prepares weights in a symmetric quantization format ahead of runtime execution.
• FP16 kernels: Direct half-precision floating point kernels are now available, expanding beyond the prior release's focus areas for mixed-precision inference.

Why it matters

For PyTorch and deep learning practitioners on AMD Zen hardware, ZenDNN 6.1 provides an alternative path to oneDNN for inference optimization while AMD and Intel's collaboration matures. The continued parallel development signals that AMD is hedging its bets, ensuring Zen-specific optimizations remain available even as industry standardization around oneDNN progresses through the UXL Foundation. Linux distribution maintainers packaging ML inference stacks will need to track both libraries to offer optimal Zen performance.