The rise of smart embedded systems marks a significant leap in technology, with uses extending from environmental monitoring to human-machine interaction. AI-driven signal processing algorithms are not only paving the way for new products, but also enhancing existing applications. As these devices become increasingly integrated into our daily lives, the pressing issue of power consumption in battery-powered devices demands innovative energy-efficient solutions.
This talk will delve into the workflow and toolchains in developing energy-efficient, high-performance AI algorithms, with a focus on optimizing signal processing for microcontrollers. We will investigate which aspects of the algorithm — whether memory operations or computations — are the most energy-intensive, and discuss architectures that are efficient in energy use. We will cover optimization techniques like quantization, which changes computations from floating-point to integer values, and graph lowering, which transforms computation graphs into symbolic representations. These techniques are crucial for optimizing both control flow and compiler levels for low-power devices like Cortex-M controllers, facilitating neural network predictions with minimal energy consumption. We will also examine the differences between software environments used for neural networks in embedded systems and those used in GPU-based training settings, underlining the limitations in computational operators and the necessary trade-offs for operations not inherently supported by microcontrollers. The talk will further highlight current efforts in deploying neural network architectures on microcontrollers.
Emerging methods that offer automatic, device-specific optimizations are making it easier to deploy AI algorithms in embedded systems more efficiently. Considering energy consumption during the design phase of these algorithms is essential. By tackling these challenges, we aim to unlock new applications and expand the possibilities within the fields of embedded systems and signal processing.