ADELIA: Analog technology creates efficient AI accelerator

The Analog Deep Learning Inference Accelerator (ADELIA) Gen2 ASIC allows rapid and energy-efficient computing of deep neural networks (DNNs).

It is one of the first modern and scalable analog AI accelerators and actually still makes use of analog computing. Having emerged from the ANDANTE project, the ADELIA Gen2 accelerator demonstrates that voice activity detection is possible with less than 200 µW of power in various applications – without compromising on accuracy.

© Adobe Stock / – / Bearbeitung Fraunhofer IIS

In many self-sufficient sensor devices, the key factor is mobility or a compact construction. Often, the energy reserves of a battery or "harvesting" systems – which collect energy from their environment – are insufficient to allow local data evaluation in the microprocessor itself. An elaborate process is therefore needed to transfer the data wirelessly to computing clusters for evaluation. "AI accelerators" such as ADELIA take a different approach and allow local data evaluation directly on the chip, thereby saving energy during transmission, relieving the burden on the wireless communication channel, and paving the way for results-based system solutions with low latency. Accelerators such as this also improve data security because, in cases of doubt, the raw data never even leave the sensor node. Researchers refer to this technology as "edge AI," whereby intelligent data analyses are carried out directly on the chip.

Edge AI applications such as this have huge potential, and the Internet of Things (IoT) is a booming, billion-dollar market. In 2028, the number of connected IoT devices is projected to reach 45 billion devices worldwide. This is associated with a massive – and steadily growing – volume of wireless data traffic. To relieve the burden, it will be necessary for constant data evaluation by AI-based sensor technology to operate more energy-efficiently than regular or query-based data transfer.

Data evaluation is the responsibility of "inference accelerators," in which the term "inference" refers to the fact that the accelerators provide the basis for the AI’s decisions. Conventional devices on smartphones use digital inference accelerators. This technology can often be found in "neural engines" in smartphone processors that use machine learning processes.

ADELIA takes this one step further, combining analog and digital computing for inference generation that is 10 times more efficient than with purely digital accelerators.

ADELIA – higher energy efficiency and accuracy with lower latency

ADELIA stands for Analog Deep Learning Inference Accelerator and offers significant advantages. This joint development by the Fraunhofer Institute for Electronic Microsystems and Solid State Technologies EMFT and Fraunhofer IIS requires just 200 μW of power to identify human speech in an audio signal. The first application of the new accelerator – a neural network for voice activity detection – achieved an accuracy of 84 percent. This is just four percentage points lower than on a computer, but ADELIA Gen2 achieves it using a fraction of the power.

The researchers have already been awarded a prize for the analysis of ECG signals using a digital AI accelerator in medicine. Now, ADELIA performs this task even more efficiently and could one day help to identify atrial fibrillation using smartwatches. "The less power we use for the subsystems, the smaller the batteries we need or the longer battery life the end products can offer," explains Dr. Markus Eppel, Group Manager Advanced Analog Circuits. In turn, this drives greater sustainability.

As well as using analog computing, it is also necessary to design specialized software. The team therefore coordinated hardware and software development in a co-design process. Unlike in digital computing, the results in analog computing are not always exact and identical. Eppel gives an example: "The calculation 3 x 2 might give you an answer of 6.2 or 5.9." The new software can get around this problem and offers highly accurate inferences without losses . The next step is to use ADELIA Gen2 to perform the keyword detection of common voice assistant systems. This "keyword spotting" process will be familiar to many users from wake words such as "Hey Siri" or "Alexa."

Great potential for the future

ADELIA Gen2 ASIC emerged from the ECSEL JU project ANDANTE, which incorporates a total of 10 partners and has a funding volume of 40.58 million euros. Following successful testing, the technology will now undergo further development, in which the team hopes to make the accelerator more powerful and energy efficient as well as trying out new use cases. Specifically, the neural network paves the way for wide-ranging applications: As well as the analysis of ECGs, Eppel also raises the prospect of object and environmental recognition by means of image evaluation. The researcher is confident that the analog system will open up countless future possibilities.

ADELIA: Analog technology creates efficient AI accelerator

Privacy warning

With the click on the play button an external video from is loaded and started. Your data is possible transferred and stored to third party. Do not start the video if you disagree. Find more about the youtube privacy statement under the following link:

Contact us

Markus Eppel

Contact Press / Media

Dr.-Ing. Markus Eppel


Fraunhofer IIS
Am Wolfsmantel 33
91058  Erlangen

Phone +49 9131 776-4415

Further topics from this research area - 2023

Bavarian Chip Design Center (BCDC)

Magazine series: Chip design in Europe