Glossary

The glossary defines the key terms used throughout the documentation.

Axon command buffer

The fundamental unit of work for Axon. A command buffer comprises compiled Axon code and a small amount of meta data.

Axon compiled model

The Axon NN compiler toolchain produces a header file nrf_axon_model_<model_name>_.h that contains the compiled model. The compiled model is entirely contained by the nrf_axon_nn_compiled_model_s structure.

Axon interlayer buffer

A global buffer shared by all Axon models for storing intermediate results. It is declared as nrf_axon_interlayer_buffer[NRF_AXON_INTERLAYER_BUFFER_SIZE] in the platform code.

Axon intrinsic

A wrapper function around a small command buffer stored in RAM, allowing certain parameters to be changed at run time. It encapsulates a specific piece of functionality.

Inference

The act of executing the machine learning model.

Packed buffer

In the context of Axon, a packed buffer with no gaps between the start of rows in a multi-dimensional shape. Axon hardware writes the start of each row to a 32-bit aligned address, so if the row size is not a multiple of 32 bits, a gap appears between rows. The compiled model structure describes the shape of the inputs and output.

Streaming model

A streaming model uses the TFLite operators VarHandle/AssignVariable/ReadVariable to save/retrieve intermediate results for the next inference. The compiled model saves these results to dedicated buffers outside of the interlayer buffer to preserve them between inferences.