ITEA 4 is the Eureka Cluster on software innovation
ITEA 4 is the Eureka Cluster on software innovation


Energy Efficient Heterogeneous AI-Framework for Smart Mobile and Embedded Systems

Project description

Essentially all mobile applications are severely power limited, which blocks huge business cases. Increasing functional complexity in mobile and autonomous applications impacts the computational load by increasing power demands of embedded platforms, making them comparable to actuation power demands. Today, it is generally acknowledged that “More AI necessitates less energy consumption”. For instance, in automotive, the present SAE-L2 of autonomy generates a significant amount of computation load in the range of approx. 1(T)FLOPS, which leads to high energy consumption. It is estimated that 5 hours of computational tasks with current e-vehicles at SAE L3 would require the energy of 100km of driving range. Computational load is a key limiter for many AI-driven concepts in smart mobility and embedded applications, which are widely emerging as functional enablers. Various technology solutions (CPU, GPU) and dedicated accelerators (FPGA, ASIC, NNP) currently compete to conquer their market domains in this regard. The trends support higher parallelism, scalability and fast memory access to optimize performance and power consumption yet lead to dedicated solutions with special market participation. EFICAS follows another path by introducing a software framework supporting energy efficient deployment of AI algorithms on the multicore heterogeneous computation technologies. It supports technology solutions, including localized and distributed computation settings. It addresses resource allocation at runtime and hybrid coherent operation with optimized task allocation at design time. For both, the EFICAS Framework targets significant improvements in the energy efficiency of AI solutions, enabling their dissemination in embedded systems in mobility, communication and automation industries. EFICAS is an AI-driven software framework, merging and utilizing heterogeneous technologies by implementing a cognitive power sensitive middleware at runtime that utilizes the performance and consumption markers of various computation technologies. EFICAS Framework targets execution of applications at a context-based optimal trade-off between performance and energy consumption by dynamically assigning specific tasks to the best-suited architecture subject to timing and geographical constraints.

Project publications