Released last November following the AMD [1]5th Gen EPYC "Turin" server processor launch was [2]ZenDNN 5.0 as their deep neural network library optimized for EPYC/Ryzen processors. ZenDNN 5.0 is their updated version of their neural network library that is compatible with the APIs from Intel oneDNN/DNNL and in turn can be used with the likes of PyTorch. It turns out ZenDNN 5.0 is capable of delivering a 400% performance uplift over their prior ZenDNN software release on the same hardware.

While writing about ZenDNN 5.0 back in November the day it was released and then forgetting about it alongside all of the other exciting Linux benchmarking and testing like their [3]AOCC 5.0 compiler , earlier this month AMD put out a blog post formally announcing the ZenDNN 5.0 software library. Not only do they talk up its Zen 5 / Turin CPU support but that ZenDNN 5.0 can provide a 400% performance uplift on average.

Earlier this month AMD engineer Shailen Sobhee outlined the ZenDNN 5.0 release and its ability to provide a 400% performance uplift. Testing across a range of models like Llama 2/3.1 compared to ZenDNN 4.2 yielded a 400% performance uplift on average. This was measured with the ZenDNN plug-in for PyTorch. There were also very nice gains over IPEX 2.4 as Intel's Extension for PyTorch.

In addition to enabling 5th Gen AMD EPYC "Turin" / Zen 5 CPU support, ZenDNN 5.0 delivers on advanced auto-tuning for LLMs, INT4 weight-only quantization, new APIs for generative LLMs, and other optimizations.

More details on the ZenDNN 5.0 software release and its 400% performance uplift via the technical article on [4]AMD Developer Central . I'll be working on some of my own ZenDNN 5.0 independent benchmarks as soon as time allows.



[1] https://www.phoronix.com/search/5th+Gen+EPYC

[2] https://www.phoronix.com/news/AMD-ZenDNN-5.0-Released

[3] https://www.phoronix.com/review/amd-aocc-5

[4] https://www.amd.com/en/developer/resources/technical-articles/zendnn-5-0-supercharge-ai-on-amd-epyc-server-cpus.html