[DATE'26]APEX: Integer-only Non-linear Function Approximation for Efficient Cross-Modal Inference

The non-linear functions introduced to modern Trans-formers are crucial to enhance the model performance. However, their high numerical precision requirements pose significant challenges for efficient inference, especially on resource-constrained hardware. Existing approximation methods still suffer from considerable computational overhead and limited generalization due to their sensitivity to the statistical distribution of activations. This limitation becomes particularly pronounced when a non-linear approximation designed for one modality is directly applied to another, as it fails to accommodate their divergent activation behaviors. To overcome these issues, we propose APEX, an efficient integer-only non-linear approximation method designed for robustness and general applicability. APEX integrates the computational graphs of non-linear functions into a unified dataflow, and performs bit-level optimization through static bit allocation and adaptive bit-width pruning (ABP), a technique that dynamically adjusts operand precision on-chip to lower computation costs and prevent underflow. We further co-design a unified and adaptive hardware architecture that supports the above two operand bit-width reduction schemes, significantly reducing hardware cost while maintaining accuracy. Extensive experiments across diverse language and vision models demonstrate that APEX achieves state-of-the-art performance. It improves accuracy by up to 0.7% on language models and 1.3% on vision models compared to prior work, with only a slight accuracy loss in multi- modal models. Furthermore, our proposed architecture achieves improvements of 1.73-8.71× in area efficiency and 1.21-10.83× in power efficiency, respectively.