AMD today released ZenDNN 5.2 as the latest versio nof their deep nueral network library that now introduces their next-generation runtime architecture. ZenDNN 5.2 is designed to deliver better performance and geater scalability over earlier versions of this AMD library that began as their take on Intel’s open-source oneDNN.
ZenDNN 5.2’s major redesign is described as “a fully re-engineered internal design offering significant gains in performance and extensibility, with full backward compatibility.” ZenDNN 5.2 now supports multiple back-ends including the native ZenDNN as well as AOCL-DLP, oneDNN, FBDGEMM, and libxsmm.
Those wanting to learn more about the major redesign with AMD ZenDNN 5.2 can do so via the GitHub announcement.
When seeing AMD ZenDNN 5.2 released, I went to see if there was any new releases of AOCC: AMD’s Optimizing C/C++ Compiler (and Fortran) for Zen processors. Sure enough, there ended up being AOCC 5.1 released at the beginning of January. No one had pointed out the AMD AOCC 5.1 release to me and completely flew under my radar.
AOCC 5.1 does bring a new Zen 5 tuned AOCL-LibM 5.2 AMD Math Library as well as various C / C++ / Fortran compiler front-end fixes. But with AOCC 5.1 it’s still based on the aging LLVM/Clang 17 release branch. LLVM 17.0 released back in September 2023. It’s sad to see the latest AOCC release still relying on this 2+ year old open-source LLVM/Clang/Flang compiler stack with all of the upstream improvements made during this time. Hopefully AMD sticking to the AOCC is a sign of them moving toward focusing on upstream LLVM and GCC compilers more. After all, Znver6 support for Zen 6 has been plumbed well in advance of these next-gen processors. Let’s hope for more upstream GCC and LLVM work from AMD this year.
For those interested the AOCC 5.1 compiler can be downloaded at AMD.com.
