Systems and methods for a multi-core optimized recurrent neural network

A non-transitory computer-readable medium or media comprising one or more sequences of instructions which, when executed by at least one processor of a computing device, causes a Recurrent Neural Network (RNN) architecture designed to operate on a processor microarchitecture to improve computing performance of the computing device, the RNN architecture comprising:

• neurons;
• modules comprising the neurons, the modules represent logical units and are arranged according to levels of a hierarchy of a processor microarchitecture to simulate a hierarchical structure of individual computing resources that comprise a processor such that each level of hierarchy is associated with at least one level of processor memory and comprises a first RNN layer and a second RNN layer, each level of processor memory further being associated with at least two of memory capacity, number of processor cores, memory bandwidth, computational bandwidth, and memory latency;
• and
• bi-directional inter-module connections that enable communication between the first and second RNN layer such as to satisfy conditions of the RNN architecture, the conditions being related to the at least two of memory capacity, number of processor cores, bandwidth, computational bandwidth, and latency.

A method for operating a Recurrent Neural Network (RNN) architecture on a processor's microarchitecture of a computing device to improve a computing performance of the computing device, the method comprising:

• based on values associated with levels of memory hierarchy that are based on a description of a processor's microarchitecture, associating each of two or more levels of a hierarchy of an RNN architecture with a level of memory hierarchy that is described by at least two of memory capacity, number of processor cores, bandwidth, computational bandwidth, and latency:
• grouping neurons into modules, each module representing a logical unit in an RNN layer within the RNN architecture;
• and
• arranging connections between the modules such that the modules satisfy conditions of the RNN architecture, the conditions being related to the at least two of memory capacity, number of processor cores, bandwidth, computational bandwidth, and latency.