Automated Search for Deep Neural Network Inference Partitioning on Embedded FPGA

Deep Neural Networks (DNNs) are currently making their way into a broad range of applications. While until recently they were mainly executed on high-performance computers, they are now also increasingly found in hardware platforms of edge applications. In order to meet the constantly changing demands, deployment of embedded Field Programmable Gate Arrays (FPGAs) is particularly suitable. Despite the tremendous advantage of high flexibility, embedded FPGAs are usually resource-constrained as they require more area than comparable Application-Specific Integrated Circuits (ASICs). Consequently, co-execution of a DNN on multiple platforms with dedicated partitioning is beneficial. Typical systems consist of FPGAs and Graphics Processing Units (GPUs). Combining the advantages of these platforms while keeping the communication overhead low is a promising way to meet the increasing requirements.
In this paper, we present an automated approach to efficiently partition DNN inference between an embedded FPGA and a GPU-based central compute platform. Our toolchain focuses on the limited hardware resources available on the embedded FPGA and the link bandwidth required to send intermediate results to the GPU. ... mehrThereby, it automatically searches for an optimal partitioning point which maximizes the hardware utilization while ensuring low bus load.
For a low-complexity DNN, we are able to identify optimal partitioning points for three different prototyping platforms. On a Xilinx ZCU104, we achieve a 50% reduction of the required link bandwidth between the FPGA and GPU compared to maximizing the number of layers executed on the embedded FPGA, while hardware utilization on the FPGA is only reduced by 7.88% and 6.38%, respectively, depending on the use of DSPs and BRAMs on the FPGA.