AMD ZenDNN 6.0 Brings Many Improvements For Accelerating Inference On Ryzen/EPYC CPUs

AMD ZenDNN 6.0 released today as a significant update to this open-source deep neural network library for helping to accelerate inferencing on AMD Zen processors from Ryzen to EPYC.

ZenDNN is AMD's alternative to the Intel / UXL Foundation oneDNN library and to serve as an AMD CPU optimized solution that can work with the likes of TensorFlow and PyTorch via plug-ins. While typically we don't begin seeing new AMD software releases until at/after new AMD CPU launches, this time around with ZenDNN 6.0 we are seeing it arrive prior to the official AMD Zen 6 debut.

With ZenDNN 6.0 there is more work on its Low Overhead API "LowOHA", expanded MoE group GEMM and FP16 operator coverage, and post-op and weight caching.

Downloads and more details on the major AMD ZenDNN 6.0 release via GitHub.