"...matrix operations with FP16 inputs and FP32 accumulation..." http://images.nvidia.com/content/volta-architecture/pdf/Volt...