Glossary - Qualcomm Dragonwing Documentation

Term	Description
AI	Artificial Intelligence
AIMET	Qualcomm AI Model Efficiency Toolkit. Tools for improving quantized model accuracy, including PTQ and QAT workflows.
API	Application Programming Interface
ASR	Automatic Speech Recognition
Backend	The execution target selected by a runtime for inference, such as CPU, GPU, or NPU/HTP.
CPU	Central Processing Unit. A general-purpose processor used for application logic and, when needed, AI inference.
CSV	Comma-Separated Value
Delegate	A runtime plug-in that offloads supported model operations from a framework (for example, LiteRT) to a hardware accelerator such as the NPU or GPU.
DL	Deep Learning
DLC	Deep Learning Container
DSP	Digital Signal Processor
GenAI	Generative AI workloads (for example, text/image generation) typically powered by LLMs or multimodal models.
GPU	Graphics Processing Unit. A parallel processor used for selected AI operations and for image/video pre- and post-processing tasks.
HMX	Hexagon Matrix Extensions
HTA	Hexagon Tensor Accelerator
HTP	Hexagon Tensor Processor. Qualcomm’s AI-optimized compute block used as the primary NPU target for efficient inference.
IM SDK	Qualcomm Intelligent Multimedia SDK for camera/video AI pipelines with zero-copy buffers and GPU-accelerated pre/post-processing.
Inference	Running a trained model to generate predictions or outputs from new input data.
IOT	Internet Of Things
LiteRT	A lightweight runtime and model format for on-device inference, often used with delegates for hardware acceleration.
Llama.cpp	C/C++ library and CLI for running LLMs and VLMs locally on CPU/GPU using quantized GGUF models.
LLM	Large Language Model. A generative model trained on large text corpora for tasks such as chat, summarization, and coding assistance.
ML	Machine Learning
Multimodal	Models that process or generate across multiple data types, such as text + image or text + audio.
NDK	Native Development Kit
NN	Neural Network
NPU	Neural Processing Unit. A specialized accelerator for running AI inference efficiently at low power (on Qualcomm platforms, typically the Hexagon HTP).
ONNX	Open Neural Network Exchange format used to represent and exchange trained models across frameworks and tools.
OpenCL	Open Computing Language
PD	Process Domain
PPA	Personal Package Archive
Preprocessing / Post-processing	Steps done before and after model inference, such as resize, normalization, decoding outputs, filtering, or drawing overlays.
PTQ	Post-Training Quantization. Quantizing a trained model without retraining.
QAIRT	Qualcomm AI Runtime SDK. Toolchain and runtime stack used to convert, quantize, compile, and run AI models on Qualcomm platforms.
QAT	Quantization-Aware Training. Training or fine-tuning with quantization effects simulated to preserve accuracy after quantization.
QNN	Qualcomm Neural Network (a synonym for Qualcomm® AI Engine Direct)
QTI	Qualcomm Technologies Incorporated
Quantization	Converting model weights/activations from higher-precision data types (for example, FP32) to lower-precision types (such as INT8) to improve performance and reduce memory usage.
RPC	Remote Procedure Call
SDK	Software Development Kit
SNPE	Snapdragon Neural Processing Engine (a synonym for Qualcomm® Neural Processing SDK)
T2T	Text-to-Text
TF	Tensorflow
TTS	Text-to-Speech
UDO	User Defined Operations
ViT	Vision Transformer
VLM	Vision-Language Model
Voice AI SDK	Qualcomm’s on-device automatic speech recognition engine (VoiceAI ASR) that runs on the NPU.
VTCM	Very Tightly Coupled Memory
WSL	Windows Subsystem for Linux
XML	Extensible Markup Language