Neuro-symbolic AI systems, which integrate neural, symbolic, and probabilistic components, emerge as a promising approach to overcome limitations of purely neural models, such as factual inaccuracies, inefficient multi-step reasoning, and poor interpretability. However, despite their superior reasoning capabilities, existing neuro-symbolic models often suffer from significant inefficiencies, hindering real-time and scalable deployment. In this paper, we systematically analyze diverse neuro-symbolic workloads, identifying critical challenges including heterogeneous compute, low ALU utilization, irregular memory accesses, and limited parallelism in symbolic and probabilistic kernels.

To address these inefficiencies, we further propose REASON, a co-design framework dedicated to accelerating probabilistic logical reasoning in neuro-symbolic AI. \underline{At the algorithm level}, REASON employs a unified acyclic graph representation with adaptive pruning to reduce model complexity. \underline{At the hardware level}, REASON features a flexible architecture with reconfigurable support for efficient symbolic and probabilistic computations, memory layout, and compilation. \underline{At the system-level}, REASON is integrated with GPU SMs with a multi-level pipelining and programmable interface. Evaluated across six neuro-symbolic workloads on 10 benchmarks, REASON exhibits 12-50$\times$ speedup and 310-681$\times$ energy efficiency compared to desktop and edge GPUs, as benchmarked under TSMC 28nm technology node. For the first time, REASON enables real-time logical and probabilistic reasoning towards human fluid intelligence, requiring only 0.8 s per task with 6 mm$^2$ area and 2.12 W power consumption.