Efficient Inference of Neural Networks with Cooperative Integer-Only Arithmetic on a SoC FPGA for Onboard LEO Satellite Network Routing

Low Earth orbit (LEO) satellite networks require real-time routing to cope with dynamic topology variations caused by continuous orbital motion. As an alternative to conventional routing approaches, deep reinforcement learning (DRL) has recently gained attention as an effective means for optimizing routing paths. To solve routing problems modeled as a grid-based Markov decision process (grid-based MDP), DRL methods such as CNN-based Dueling DQN have been proposed. However, these approaches are difficult to implement in practice. In particular, the substantial floating-point computation and memory traffic of CNN inference make real-time onboard inference challenging under the stringent power and resource constraints of satellite platforms. To address these constraints, this paper proposes an INT8 quantization and hardware–software co-design framework using heterogeneous SoC FPGA acceleration. We offload compute-intensive CNN inference to the programmable logic (PL), while the processing system (PS) orchestrates overall control and data movement, forming a collaborative PS–PL architecture. Furthermore, we integrate the NITI-style two-pass scaling with PS–PL exponent propagation to preserve end-to-end integer consistency without floating-point conversion. To demonstrate its practical onboard feasibility, we employ standard accelerator implementation choices—such as output-stationary scheduling and on-chip prefetching—and conduct an ablation study over independently tunable axes (PE array size and PS-side buffer reuse) to quantify their incremental contributions. Experimental results show that the proposed PS–PL cooperative scheme dramatically reduces computation time compared to a PS-only reference implementation on the same platform.