UXL's oneDNN 3.13 Preps For Intel Nova Lake With AVX10.2, More Intel Optimizations

Following the release of AMD's ZenDNN 6.0 earlier this month, there is a new feature release of the oneDNN neural network library that used to be developed by Intel as part of oneAPI and is now under the UXL Foundation umbrella. Even so, oneDNN feature releases continue to be heavy on new Intel optimizations and future hardware support. The oneDNN 3.13 release improves performance for upcoming Intel Core Ultra CPUs featuring the AVX10.2 instruction set, a.k.a. Nova Lake. In addition to AVX10.2 tuning, oneDNN 3.13 also brings MATMUL performance improvements for existing Intel Xeon CPUs with the AMX instruction set. There are also some AVX2 optimizations for those CPUs without AVX-512 / AVX10 / AMX, various other CPU tunings, and also some AArch64 and RISC-V processor optimizations. On the graphics side, oneDNN 3.13 improves performance for upcoming Intel integrated Xe3P-LPG graphics to be found with Nova Lake P. There is also some convolution performance optimizations for current-generation Arc B-Series "Battlemage" graphics cards. Downloads and more details on Friday's release of oneDNN 3.13 via GitHub.