Extensible multi-precision data pipeline for computing non-linear and arithmetic functions in artificial neural networks

A hybrid computational system, comprising:

• a configurable and programmable specialty hardware programmed by a central processing unit with model data determined by a graphical processor unit to implement neural networks associated with an extensible multi-precision data pipeline;
• a memory interface controller;
• a memory;
• and
• a central interconnect;
• wherein the extensible multi-precision data pipeline, comprises:
• a local buffer that loads input local data from the memory and stores the input local data set in a local storage format of either 32-bit fixed-point, 16-bit fixed-point, or s-bit fixed-point;
• an input tensor shaper coupled to the local buffer that reads the input local data set in the local storage format in the local buffer and converts the input local data set into an input tensor data set having a 2-dimensional tensor format of vector width N by tensor length L wherein N and are integers;
• a cascaded pipeline coupled to the input tensor shaper that routes the input tensor data set through at least one function stage resulting in an output tensor data set;
• element-wise stage within the at least one function stage that processes the input tensor data set on an element by element basis along the tensor length L;
• an output:
• tensor shaper coupled to the cascaded pipeline that converts the output tensor data set into an output local data set having the local storage format;
• and
• wherein the output tensor shaper writes the output local data set to the local huller to be stored on the memory.