Rather than these, the original Winograd algorithm ("A New Algorithm for Inner Product", 1968) is more useful in practice, if not for matrix multiplication in s

Rather than these, the original Winograd algorithm ("A New Algorithm for Inner Product", 1968) is more useful in practice, if not for matrix multiplication in software on the typical CPU architecture. This decomposition is directly related to the Winograd minimal filtering algorithms for fast convolution used in convnets today. Also, systolic array structures for matrix multiplication in hardware exist that take advantage of this reformulation to reduce the number of multiplies, as the cost in hardware is substantially lower for addition versus multiplication.