Inferensys

Glossary

Always-On (AON) Domain

An Always-On (AON) Domain is a small, isolated section of a system-on-chip that remains powered to monitor sensors and handle wake-up events while the main processor sleeps.
Developer building agentic RAG system, retrieval pipeline diagram on laptop, technical workspace with notes.
POWER-AWARE TINYML

What is Always-On (AON) Domain?

An always-on (AON) domain is a small, isolated section of a system-on-chip that remains powered even when the main processor and other subsystems are in a deep sleep state, responsible for monitoring critical sensors and handling wake-up events.

An Always-On (AON) Domain is a dedicated, ultra-low-power subsystem within a System-on-Chip (SoC) or microcontroller that remains active during the device's primary sleep states. It contains a minimal set of components—such as a simple processor core, real-time clock (RTC), wake-up controller, and dedicated SRAM—powered by a separate, highly efficient voltage rail. Its primary function is to maintain basic system context and monitor for external triggers without waking the energy-intensive main compute cores, enabling always-on sensing capabilities.

In TinyML and IoT applications, the AON domain is critical for battery life extension. It runs simple, stateless inference models or threshold detection logic on data from connected sensors. When a predefined event is detected, the AON domain's controller initiates a wake-on-event sequence, powering up the main application processor only when necessary. This architecture is fundamental to achieving energy-proportional computing in edge devices, as it minimizes the leakage power and dynamic energy wasted during long idle periods.

ARCHITECTURAL OVERVIEW

Key Components of an AON Domain

An Always-On (AON) domain is a specialized, isolated subsystem within a System-on-Chip (SoC) designed for ultra-low-power monitoring. Its architecture comprises several critical hardware and software elements that enable continuous operation while the main application processor sleeps.

01

Ultra-Low-Power Processor Core

The computational heart of the AON domain is a dedicated microcontroller or processor core designed for minimal leakage current. These cores are typically:

  • RISC-V or Arm Cortex-M0/M0+ class architectures.
  • Optimized for sub-threshold or near-threshold voltage operation.
  • Lack complex features like out-of-order execution or deep pipelines to minimize active and static power.

Their sole purpose is to execute simple control logic and sensor data preprocessing with microwatt-level power budgets.

02

Always-On Memory (AON RAM/ROM)

A dedicated, power-gated memory block that remains accessible only to the AON domain. This includes:

  • Small, retained SRAM (e.g., 4-64 KB) for storing state, sensor data, and wake-up code.
  • Non-volatile memory (ROM/Flash) for storing the immutable AON firmware bootloader.
  • Memory is isolated from the main system bus to prevent leakage paths from the powered-down main memory.

This memory is crucial for maintaining context and enabling instant resume without booting the main OS.

03

Wake-Up Controller & Event Detectors

A hardware block that monitors digital and analog inputs for pre-programmed wake-up events without CPU intervention. It handles:

  • GPIO pin state changes (e.g., button press).
  • Analog comparator outputs (e.g., sensor threshold crossing).
  • Timer/RTC alarms for scheduled wake-ups.
  • Wake-up on pattern from low-power communication interfaces like UART or I2C.

Upon detecting a qualifying event, it triggers an interrupt to the AON processor or directly initiates the main power rail sequencing.

04

Low-Leakage Power Rails & Isolation

The AON domain is powered by a separate, always-on voltage regulator that is highly efficient at very low load currents. Key techniques include:

  • Power gating isolation cells at the boundary between the AON domain and the powered-off main logic to prevent leakage.
  • Level shifters for voltage translation between domains.
  • Retention flip-flops that maintain state with a tiny trickle of power while the rest of a block is shut down.

This electrical isolation is fundamental to preventing the AON domain from being a parasitic load on the battery.

05

Sensor Peripherals & Data Path

A set of minimal, low-power peripherals directly accessible by the AON processor for interfacing with critical sensors. This often includes:

  • Low-Speed I2C/SPI controllers for environmental sensors.
  • Low-power ADC for sampling analog sensors (e.g., thermistor, microphone).
  • Pulse counters for simple event counting from motion or hall-effect sensors.
  • Dedicated hardware accelerators for basic DSP tasks like FFT or voice activity detection to offload the AON CPU.

Data is pre-processed here to determine if a full system wake-up is justified.

06

AON Firmware & Real-Time OS

The software layer that defines the AON domain's intelligence. It is a deterministic, bare-metal or RTOS-based application responsible for:

  • Polling or interrupt-driven sensor management.
  • Running simple TinyML models (e.g., keyword spotting, anomaly detection).
  • Managing the wake-up logic and sequencing the main processor's power-up.
  • Communicating with the main OS via shared memory or messaging queues upon wake-up.

This firmware is typically statically linked and stored in ROM or protected flash to ensure reliability.

ARCHITECTURAL COMPARISON

AON Domain vs. Main Application Processor Domain

A comparison of the key architectural and operational characteristics of the Always-On (AON) domain and the Main Application Processor (AP) domain in a System-on-Chip (SoC) designed for power-aware TinyML.

Feature / CharacteristicAlways-On (AON) DomainMain Application Processor (AP) Domain

Primary Purpose

Ultra-low-power sensing, event monitoring, and wake-up trigger management.

High-performance application execution, including complex ML inference and user interface tasks.

Power State During System Sleep

Remains powered (ON).

Fully powered down or in deepest sleep state (OFF).

Typical Power Consumption

< 1 mW

10 mW to 1 W+

Core Components

Ultra-low-power processor (e.g., Cortex-M0+), always-on memory (SRAM/retention RAM), wake-up controller, simple peripherals (GPIO, low-speed ADC).

High-performance cores (e.g., Cortex-A series), large caches, main system memory (DRAM), complex peripherals (GPU, NPU, high-speed interfaces).

Wake-up Capability

Can wake the main AP domain via interrupt signals upon detecting a predefined sensor event or timer expiration.

Woken by external interrupts, typically from the AON domain. Has high wake-up latency.

Memory Footprint & Type

KB-range of tightly coupled, low-leakage SRAM (retention RAM).

MB to GB range of volatile DRAM and large caches.

Clock Frequency

KHz to low MHz range (e.g., 32 kHz - 50 MHz).

High MHz to multi-GHz range.

Typical Workloads

Simple threshold detection on sensor data (e.g., keyword spotting, motion detection), real-time clock (RTC) management, GPIO toggling.

Running full operating system, executing large neural networks, processing camera frames, handling network stacks.

Direct Sensor Interface

Independent Power Rail

Software Complexity

Bare-metal or RTOS-based firmware.

Full-featured OS (Linux, Android) with complex application stacks.

Design Priority

Minimize leakage current and dynamic power at all costs.

Maximize performance and throughput within a thermal/power envelope.

POWER-AWARE TINYML

Frequently Asked Questions

An always-on (AON) domain is a critical hardware design pattern for ultra-low-power edge AI and IoT devices. It enables continuous sensor monitoring and instant system wake-up while the main processor sleeps, drastically extending battery life. This FAQ addresses its core mechanisms, design considerations, and role in TinyML systems.

An always-on (AON) domain is a small, isolated power and clock domain within a System-on-Chip (SoC) or microcontroller that remains fully powered and operational even when the main processor cores, memory, and most peripherals are in a deep sleep or powered-off state. Its primary function is to monitor critical wake-up events from sensors or timers with minimal energy expenditure, enabling the rest of the system to remain dormant until needed. This architecture is foundational for battery-powered IoT devices and wearables that require months or years of operation from a single charge, as it eliminates the energy waste of periodically waking the entire high-power system just to check a sensor.

Prasad Kumkar

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.