DMA with In-Flight Data Transformation
- Subject:Digital Hardware Development, Machine Learning
- Type:Masterarbeit
- Date:ab 09 / 2025
- Tutor:
DMA with In-Flight Data Transformation
Context
The execution of Deep Neural Network (DNN) inferences for machine learning applications in embedded Multi Processor System-on-Chips (MPSoCs) often results in high memory consumption and substantial traffic through the on-chip network. To further optimize the throughput of such platforms it would be beneficial to relieve the processor cores, by offloading the computations of repetetive data transformations. For that purpose we want to use an extended DMA module, which is able to apply common data transformations during transfers.
Tasks
The work includes the following tasks:
- Survey of state of the art DMA functionalities
- Identifying typical data transformations in CNNs
- Development and Implementation of an extended DMA module
- Evaluation of the extended DMA in Simulation and on an FPGA
Requirements
- Knowledge of a hardware description language: (System)Verilog or VHDL
- Knowledge in the field of hardware design testing and verification
- Experience working with Python
- Ideally, knowledge of SoC architectures