To achieve these performance increases, KNM is introducing new instruction sets that improve low-precision computing called Quad Fused Multiply Add(QFMA) and Quad Virtual Neural Network Instruction (QVNNI). The QFMA can up to double the amount of single precision performance KNM can deliver over that of the current Intel Xeon Phi processor and QVNNI, which reduces the precision further, bringing additional performance gains for deep learning workloads. With both QFMA and QVNNI, Knights Mill can deliver significantly higher training performance than today’s Intel Xeon Phi processor, making it a targeted solution for deep learning workloads. The new Knights Mill is expected for production in the fourth quarter of 2017.
QVNNIW, QFMAPS
AVX-512 Vector Neural Network Instructions Word variable precision (QVNNIW) - vector instructions for deep learning, enhanced word, variable precision.
AVX-512 Fused Multiply Accumulation Packed Single precision (QFMAPS) - vector instructions for deep learning, floating point, single precision.