A Compiler-Aware Framework for Partitioned Neural Network Inference on FPGA DPUs

The increasing adoption of Artificial Intelligence is driving the development of larger and more accurate neu-ral networks. However, their high computational cost leads to significant inference latency, especially on resource-constrained embedded platforms such as FPGA-based systems. To address this issue, AMD introduced the Deep Learning Processing Unit, an accelerator integrated into the Vitis AI toolchain for efficient execution of quantized neural networks. Nevertheless, models with many parameters can be difficult to deploy on a single board due to latency or capacity constraints. In these scenarios, splitting a model into separately compilable fragments becomes useful for more flexible deployment. This process is not immediate within the Vitis AI flow. Our analysis shows that manually partitioning a network into independently compiled sub-networks can remove compiler optimizations. In particular, losing the global view of the quantized XIR graph can prevent the toolchain from preserving DPU mapping, moving accelerable operations to the CPU and causing severe performance degradation. Based on this observation, we analyze the Vitis AI compiler and propose an XIR-level splitting framework that generates independently compilable .xmodel fragments while preserving the context required for DPU mapping. The approach keeps the workflow high-level, without requiring advanced hardware design expertise or low-level design changes. Experimental results on CNN and ConvViT-based models show that naive splitting can introduce slowdowns up to ×249 on CNNs and ×542 on ConvViT variants. The proposed framework restores correct DPU mapping by addressing boundary-context loss and incomplete dependency collection, bringing latency back to the expected range for hardware-accelerated execution.