LODRIC is the successor project of „LO3-ML“ which earned a 1st Prize in a Germany-innovation contest for its high energy efficiency.
The goal of „LODRIC“ is the development of design methodologies for low-power digital AI-circuits with embedded non-volatile memory cells (ReRAM). The main innovations from the „LO3-ML“ project, namely the design of a dataflow oriented computer architecture in connection with distributed non-volatile memory for the weights of the neural network and ternary quantisation of weights, should be revisited and evolved.
The developed architecture for the interference of neural networks is to be prototypically tested using three different applications: analysis of ECG data, predictive maintenance and radar signal processing.
The reasoning behind using a custom architecture is to reduce the overhead coherent in general-purpose architectures und therefore enable more efficient calculations. For the interference of neural networks systolic arrays are a promissing architecture. The integration of ReRAMs on the same chip as the logic circuit enables to exclude external memory and therefore the energy cost involved with the communication overhead. Further the use of ternary logic (-1, 0 , 1) significantly reduces hardware complexity, especially for multiplications, which also reduces overall power consumption.
The project takes place in collaboration with the Technical University of Munich (Chair of Micro- and Nanosystems Technology) and the Fraunhofer Institute for Integrated Circuits (departements IIS-KS, IIS-EAS and IIS-SSE).