+33 (0)6 85 22 07 25 mario.diaznava@st.com

ANDANTE RESULTS AND VALUE CHAIN

This Newsletter presents the results obtained in the ANDANTE project financed by the ECSEL/H2020 program.

ANDANTE aims to leverage innovative integrated circuit designs to build powerful hardware and software components and devices for artificial and edge neural networks, which will serve as a starting point for future edge products, combining extreme power efficiency with robust neuromorphic computing capabilities. Specifically, the hardware and software components to be deployed in the micro-edge, where the data is generated. (See figure below).

Edge AI Architecture

The main requirements of micro-edge are:

  • Data diversity: the data source can come from many different types of sensors closely related to the application.
  • Autonomy: Extremely low power consumption is necessary since most of these devices are battery operated.
  • Low latency: Highly efficient data processing to ensure latency requirements.
  • Accuracy: AI models must be the most appropriate to meet this requirement.
  • Cost: This is an important factor in commercial acceptance.
  • Others not addressed by ANDANTE such as security, reliability, etc.

Existing solutions cannot meet all these requirements given the wide variety of applications. This then requires the design of new components, and their implantation must also make a compromise between technology and cost without forgetting main application requirements. ANDANTE’s design results, although not a surprise, show that dedicated ASICs solutions are highly efficient (e.g. power consumption in the range of mW and few mW) at the expense of flexibility. SoCs with multiple processing units are more flexible and different models can be implemented, but the price is in energy consumption, which can reach several hundred mW. Finally, the FPGA has very high programmability but a poor power/performance ratio, reaching tens of Watts.

ANDANTE covered the completed value chain, as shown in the Figure below. It includes the following activities: Development of novel eNVM technologies to reduce energy consumption by reducing the data movement; tools to facilitate the good design of neuromorphic processors; the design of neuromorphic processors implementing artificial or spiking neural networks and the associated platforms allowing the exploitation of these devices; finally, the fourteen use cases defined in five application areas to evaluate all the components (HW and SW) developed in ANDANTE.

Edge AI Architecture

The following sections describe the activities described in ANDANTE in more detail.