Execution Time Estimation of ML algorithms on PLCs

Executing numerical algorithms on microcontrollers is difficult and usually an engineering project on its own. If you are faced with many potential algorithms (as is typically the case when you consider solving a task using machine learning), it is vital to establish a solid understanding of what algorithms to invest in - within limited budget and time. We make this possible.

We maintain a MLIR-based compiler tool chain allowing for mapping of trained ML models (ONNX, sklearn, pytorch, Tensorflow) onto PLCs and microcontrollers. Currently we target ARMv7 (i.e. Cortex M4) as well as ARMv8 ISA (i.e. Cortex-A72), in the future also RISC-V. Our tool chain produces executable code, thereby allowing for realistic run-time and power consumption estimates of complex algorithms on the shop floor. Our report can be used to steer model development in a PLC-friendly direction or evaluate the risk of model development activities. We constantly monitor newest software releases and can advise on supported ML methods in ONNX, sklearn,pytorch, TF and MLIR.