Abstract: Systems and methods include technology that divides first data associated with a first matrix into first multiplication tiles based on a block size that is identified based on an available space of a memory of an accelerator, and second data associated with a second matrix into second multiplication tiles based on the block size. The technology divides the first multiplication tiles into a plurality of first groups that correspond to a plurality of matrix multiplication operations and the second multiplication tiles into a plurality of second groups that correspond to the plurality of matrix multiplication operations. The technology loads a selected first multiplication tile of the first multiplication tiles and a selected second multiplication tile of the second multiplication tiles into the memory to execute one or more of the plurality of matrix multiplication operations with selected groups of the first plurality of groups and the second plurality of groups.

Abstract: Graph reordering and tiling techniques are described herein. In one example, large graphs (e.g., for inferencing with graph neural networks) can be reordered, tiled, or both, to achieve maximal data reuse and uniform compute load distribution. In one example, a reordering method involves performing breadth first search (BFS) renumbering on a graph data set with the highest degree destination node as the root node to generate a reordered graph data set. BFS is then performed again with candidate nodes from the last level of the reordered graph. The second reordered graph data set with the lowest bandwidth or best profile can be selected for further processing. In one example, a method of tiling involves dividing a graph data set into tiles to balance expected compute time.