Hardt, WolframSaleh, ShadiBeyer, VolkhardKoppala, Monisha RaoMonisha RaoKoppala2025-05-082025-05-082025-04-29https://publica.fraunhofer.de/handle/publica/48744510.60687/2025-0048In this thesis, the Row Stationary (RS) dataflow introduced in Eyeriss is suggested for the MCU. RS maximizes the input reuse and minimizes the partial sum movement. This reduces the energy consumption of the model. The LeNet model is chosen to employ different data types reuse of RS dataflow onto the MCU. The convolutional layers are implemented with input pixel (row-wise) and filter weight reuse (row-wise) and fully connected layers are implemented with input pixel and psum reuse. An analysis framework is considered to compare the conventional and RS dataflow under the same memory area. The CONV layers of RS dataflow save energy and latency up to 14.42 % and 14.42 % respectively in comparison to the conventional dataflow. On the other hand, FC layers save both energy and latency up to 16 %. The energy and latency of the LeNet model are improved by almost 14 % by reducing the SRAM memory accesses for RS dataflow.enCNNMCURow StationaryEnergyLatencyConvolutional Neural NetworkMicrocontroller000 Informatik, Informationswissenschaft, allgemeine Werkemaster thesis