Memory Allocation for Neural Networks using Graph Coloring
Leonid Barenboim, Rami Drucker, Oleg Zatulovsky, Eli Levi · 2022
The Memory Allocation problem for neural networks can be represented as a two-dimensional optimization problem. The neural network is allocated into limited memory space while allocating as much data as possible into the low latency memory. Our solution is based on a generalization of graph coloring, edge-to-node transformation and considers the order in which the graph nodes are colored. We observed improvement of more than 40% in SRAM memory bandwidth in various neural networks.