From what i can quickly gather from the article:
This is all based on magnonics, which in short - is the use of magnetic spin for binary storage and or logic. This device focuses on the later...
From what I could tell from the article, it appears to focus on both. The device allows them to quickly create and store a pattern. It then allows another pattern to be created and quickly compare the patterns.
It does this by constructing a matrix of magnetic nodes that are effectively interconnected to neighbours (moor?) via spatial magnetic-spin sensitivity, these interconnects form the dynamic logic processing ability of the matrix.
I think that this is somewhat like a (soft) convolutional artificial neural network for image recognition, these are constructed out of a 2d or 3d matrix of nodes with weighted interconnects in a moor-neighbourhood arrangement. The difference here i guess is that a) it's done with magnetic spin (i really have no idea why this is an advantage, maybe i'm all wrong about this) and b) being an application specific piece of hardware each node works in parallel (this is trumped as the primary reason for the speed potential in the article).
... Big disclaimer: I am massively speculating because the use case is not made super clear.
From what I could tell, the advantage of the magnetic spin is that with an 8 terminal node, it can quickly create a matrix of 1,000 bits of data. It can then compare that matrix with a stored matrix. With several nodes working in parallel, it takes the same time to compare 1,000 bits of data as 10,000,000. The potential (and right now, only potential) is that a computing device with this hardware integrated in could compare an image with a stored image much faster than conventional circuits could make the comparison and result in a better matching algorithm.