Intel MKL-DNN / DNNL 1.3 Released With Cooper Lake Optimizations

Intel on Thursday released version 1.3 of their Deep Neural Network Library (DNNL) formerly known as MKL-DNN in offering a open-source performance library for deep learning applications.

Notable to DNNL 1.3 are "broad release quality optimizations" for upcoming Intel Xeon "Cooper Lake" processors. The Cooper Lake optimizations are there, granted, it was recently revealed that Intel is only going to be offering Cooper Lake for quad/octal-socket Xeon Scalable platforms. For those with just single or dual socket platform interest, it's now Cascade Lake Refresh until Ice Lake Xeon CPUs ship in future quarters. With Cooper Lake there is BFloat16 support and presumably most of the DNNL 1.3 optimizations are centered on BF16 capabilities.

Besides the Cooper Lake optimizations, DNNL 1.3 also has improved performance for its matrix multiply "matmul" primitive for 3D tensors, better binary primitives for where tensors have to be broadcasted on all processors, and better performance of convolution primitives for 3D tensors.

DNNL 1.3 also has matmul primitive support for Intel processor graphics, new filters, extending existing primitives, and other enhancements.

More details on the MKL-DNN/DNNL 1.3 changes via GitHub. I'll be having some fresh DNNL CPU benchmarks up shortly.