Glossary

DLPU — Deep Learning Processing Unit. A general name for the hardware designed to accelerate deep learning tasks.

SoC — System on Chip. In this context, the specific computing system on which the DLPU can be used. Different products can have different SoCs which behave differently.

Quantization — The process of converting high-precision weights and activations (often represented as float32), to lower-precision formats (often int8/uint8). This is to shrink the model's size and memory bandwidth.