Neural network accelerator with compact instruct set

A method, comprising:

• receiving, by a neural network processor, program code comprising a first instruction, a second instruction, and a third instruction, the first instruction including a first operational code that identifies the first instruction;
• executing, by the neural network processor, the first instruction, wherein executing the first instruction comprises:
• determining, from a base address included in the first instruction, a first location where a first set of weights are stored in a memory of the neural network processor;
• determining, from the base address, a second location where a first set of values are stored in the memory;
• retrieving the first set of weights and the first set of values from the memory;
• broadcasting the first instruction to each of a first processing unit and a second processing unit of the neural network processor;
• generating, by the first processing unit, a set of products, wherein each product from the set of products is a result of multiplying a weight from the first set of weights with a value from the first set of values;
• generating, by the second processing unit, a summed value by adding together all products from the set of products, wherein the summed value represents an output activation of a neural network;
• determining, by the neural network processor, that the second instruction is immediately followed in the program code by the third instruction from an operational code of the second instruction;
• obtaining, by the neural network processor, information from the third instruction;
• and
• executing, by the neural network processor, the second instruction using the information.