ENERGy optImised analogous Circuits for deep neural networkS (ENERGICS)

The ENERGICS project is a BMBF project in cooperation with the Chair for Cardiovascular Bioinformatics (Prof. Marcel Schulz) FB16 Medicine. It is part of the "Energy Efficient AI System" competition project BMBF project . The University of Frankfurt starts in the 130nm class.

Energy-efficient AI systems are extremely important for the future. Autonomous driving, image recognition or medical applications, diagnosis or only speech or text recognition can be tackled by trained neural networks for classification (CNNs). In any case, this frequently used solution requires a lot of energy since numerous computing operations have to be carried out. In the case of motor vehicles, this led to considerable additional energy requirements per kilometer driven, as well as cooling problems. In medical applications, battery life can also be greatly improved  by such an energy aware AI.

The project contributes in the competition "energy-efficient AI systems", which enables a given medical classification problem to be calculated with the lowest possible energy consumption. The pursued approach exploits the possibility of a direct implementation of neurons in analog hardware and therefore uses the immediately available analog representation for the signals as opposed to a digital representation in processor-based designs. The project would like to optimize the proposed architecture on several levels with the goal of minimal energy consumption.

The implementation is planned as follows.

In the ENERGICS project, a new three-stage optimization process is being developed, which allows the design of an extremely energy-saving analog circuit for a deep neural network. The method is being developed to implement the prediction of arrhythmia patterns from ECG data in an energy-efficient manner. The following strategies are implemented:

  • Optimization of analog circuits to make maximum use of the energy saving potential. Fully automatic topology synthesis and sizing of the analog chip offers the highest energy saving potential at transistor level and automatic creation of a large number of highly optimized basic blocks for neuron configurations.
  • Software optimization of the neural network.
  • Cross-level optimization enables the savings potential at system level with optimizations at the electrical level and leads to the best possible tradeoff of accuracy and energy-saving potential for the detection of arrhythmia patterns from ECG data.