ANDANTE - Ai for New Devices And Technologies at the Edge#
Project duration: 2020 - 2023
Various hardware approaches are under discussion for the implementation of neuromorphic computing. Basically, such a computer is constructed comparable to a human brain - from countless synapses and neurons. The ANDANTE project uses ferroelectric field effect transistors (FeFETs) as the ideal synapse. These are characterized by a very high dynamic range as well as very low propagation delay. Thus, the signals of the pre-neuron can be transmitted quickly and with very low loss and collected in the post-neuron. In addition, these can be incorporated into a chip alongside standard logic transistors, enabling a scalable edge AI accelerator.
"Edge AI" refers to an approach in which intelligent data analysis is performed directly on the computer chip, enabling increased computational speed while reducing power consumption. Fraunhofer IPMS is implementing FeFETs directly on the semiconductor chip to thus develop the most energy-efficient edge AI accelerators available on the market.
These chips make it possible, for example, to assign objects in images to different categories or to segment areas in an image. However, algorithms that process these tasks have a high memory access, for example to describe the connection of neurons, the synapses. To solve this problem, chips are being developed that directly embed the function of the synapse, as in the human brain. The calculations are thus realized very efficiently, since no complicated data transfer between computer and processor is necessary anymore. This enables decentralized, lightning-fast and energy-efficient data processing for sensors in the Edge. In the future, sensors can thus be evaluated decentrally without having to send their data to the cloud.
In addition, Fraunhofer IPMS is also demonstrating compatible neuron circuits in the ANDANTE project. In close cooperation with FMC and the Fraunhofer Institutes IIS and EMFT, work is being done on a Fraunhofer FeFET ASIC that can be manufactured in Globalfoundries' 22FDX® technology. The declared goal here is to develop the most energy-efficient AI chip on the market.