I’m happy to announce that our neural-network-native-ops (aka NNNOps) – a simple, yet powerful wrapper for common numerical operations typically used in deep neural networks – now supports Intel MKL as its backend.
Intel MKL fully utilizes newest CPUs (including Skylake) with all their bells and whistles (including AVX-512). In our scenario this means up to 2.5x faster inference (single-threaded 512x512 matrix multiplication on a Skylake Gold CPU).
But, as always, your mileage may vary, so go test your case – with the newest NNNOps you can easily switch backends.