Low precision matrix multiplication for efficient deep learning in NVIDIA Carmel processors

We introduce a high performance, multi-threaded realization of the gemm kernel for ARMv8.2 architecture that operates with 16-bit (half precision)/queryKindly check and confirm whether corresponding author is correctly identified. floating point operands. Our code especially designed efficient machine learning inference (and to certain extent, also training) deep neural networks. The results on NVIDIA Carmel multicore processor, which implements architecture, show considerable performance gains kernel, close theoretical peak acceleration could be expected when moving from 32-bit arithmetic/data 16-bit. Combined type convolution operator arising in convolutional networks, speed-ups are more modest though still relevant.