Fraunhofer Institute for Integrated Circuits IIS
Fraunhofer Institute for Integrated Circuits IIS

ADELIA:
Ultra-Efficient Edge AI Accelerator for Low-Power On-Device Intelligence

Low-Latency AI Directly on Devices with Ultra-Low Power Consumption

Energy-Efficient Inference of Deep Neural Networks Directly at the Edge

Photography of ADELIA: analog deep learning inference accelerator
© Adobe Stock / www.freund-foto.de – stock.adobe.com / edited by Fraunhofer IIS

Running AI at the edge remains a major challenge. Deep neural networks (DNNs) require significant compute power, leading to high energy consumption and latency. Conventional microcontrollers and digital accelerators often cannot meet the demands of modern, always-on AI applications.
ADELIA introduces a fundamentally new approach to AI acceleration. As a mixed-signal AI accelerator, it combines analog and digital computing to execute deep neural networks with exceptional energy efficiency.

At its core are configurable LP‑vxCores that perform key neural network operations using analog computing principles, complemented by flexible digital control logic. This hybrid architecture enables highly parallel execution of multiply-and-accumulate operations – the computational backbone of AI – while drastically reducing power consumption compared to purely digital approaches.

The result is efficient, real-time AI processing directly where data is generated: at the edge.

Proven Edge AI Performance:
On-Device Intelligence with Neuromorphic Audio Processing

ADELIA supports keyword spotting with 10 classes at an accuracy of approximately 92%, accessing the raw audio data directly. On comparable semiconductor nodes, the implementation achieves up to 97 times lower power dissipation and 41 times faster inference time, with inference latency evaluated without input current and effective throughput evaluated including input current.

ADELIA supports keyword spotting with 10 classes at an accuracy of 92%, accessing the raw audio data directly. On comparable semiconductor nodes, the implementation achieves up to 97 times lower power dissipation and 41 times faster inference time, with inference latency evaluated without input current and effective throughput evaluated including input current.

Low-Power Edge AI Features and Performance Benefits

ADELIA Key Facts ADELIA Benefits
  • Configurable mixed-signal LP-vxCores
  • Scalable multi-core architecture on system level
  • Highly parallel MAC execution
  • Support of established semiconductor processes, e.g. GlobalFoundries 22FDX®, TSMC 90nm
  • Hardware-software co-design flow
  • Available as standard or custom IP core
  • Available as evaluation kit
  • Up to 90% lower power consumption compared to digital accelerators
  • Fast reaction time because of massively parallel computation
  • Versatile deployment of various neural networks
  • Adapts easily to various performance needs
  • Early and efficient performance verification
  • Easy integration into existing designs
  • Fast start and quick prototyping

Flexible Edge AI Solutions for Efficient On-Device Deployment

The ADELIA technology enables efficient deployment of deep neural networks across a wide range of edge applications. Our expertise spans AI model development, hardware design, and system integration, supported by a specialized co-design flow that reduces time-to-market. Based on this end-to-end expertise, we offer a flexible portfolio of solutions from ready-to-use accelerators to customizable IP and dedicated development services:

Edge AI Accelerators
  • Complete accelerator designs ready for production
  • Custom designs tailored to the application
LP-vxCore
  • Accelerator cores available as standard or custom designs
  • Integration and adaption to customer circuit designs
Services
  • Feasibility studies
  • AI model development and deployment
  • Workshops
  • Support for productization via partner network

Industrial Applications of Ultra-Low Power Edge AI

 

© Fabian – stock.adobe.com /Fraunhofer IIS

Download

Predictive Maintenance: Intelligent Bolts with Embedded Edge AI

Manual inspections are no longer enough. With the embedded ADELIA AI chip, the NeuroQ-Bolt enables continuous structural health monitoring and identifies damage before it becomes visible.

 

© adobestock.com / Shuo / Composing by Fraunhofer IIS

Download

Digital Health Monitoring: Real-Time Cardiac Sensing with Edge AI

Powered by ADELIA, neuromorphic computing brings real-time ECG analysis to wearables enabling energy-efficient detection of cardiac arrhythmias directly on the device.

© panuwat - stock.adobe.com

Upcoming

Embedded Machine Vision: Low-Power Image Recognition at the Edge

Already under development is the next generation of ADELIA – a scaled up SoC containing LP-vxCores and a RISC-V® MCU with massively improved throughput and memory capacity. With this chip-version, image classification in edge applications using the MobileNetV2 model will be possible with extremely low power consumption.

This might also interest you

Deep Dive

Our Selected Scientific Publications and Talks on Ultra-Efficient Edge AI

 

© Fraunhofer IIS

Press release / 3.3.2026

AI Chip Design From Heilbronn

Fraunhofer HNFIZ adds integrated circuits for and with AI to its service range

Read more

Press release / 12.12.2019

Germany’s Sprunginnovationen competition with the ADELIA project

To the press release

ADELIA: Analog technology creates efficient AI accelerator

Everything digital now? Not at all! ADELIA brings analog computing back – enabling fast, energy-efficient AI.

To the article
 

© Shuo-stock.adobe.com

A strategic initiative of Fraunhofer IIS

Neuromorphic Computing

Learn how neuromorphic computing at Fraunhofer IIS contributes to efficiency improvements in various applications.

More information
 

© Adobe Stock / Sergey Tarasov - stock.adobe.com

SNNs – Spiking neurons with a sense of time

Spiking Neural Networks (SNNs) are a subset of neuromorphic computing. Therefore, we not only explore their topology but also develop customized hardware designs for ultra-low-power and ultra-low-latency applications – making Artificial Intelligence even more similar to the human brain.

Spiking Neural Networks
 

© Fraunhofer IAF

Fraunhofer Heilbronn Research and Innovation Centers HNFIZ

Research and Innovation Center for Chip AI

To FIZ Chip AI
 

© Fraunhofer IIS/Bianca Möller

Neuromorphic
Computing Labs of Northern Bavaria

»Neuromorphic Computing Labs of Northern Bavaria« is an association of basic and applied research that focuses on the complex topic of neuromorphic engineering. Close cooperation between leading institutions in Northern Bavaria is intended to drive forward innovative solutions. In addition, a series of lectures provides regular insights into current research.

NeuroComp Labs NBay