<p>Post-stack migrated 3D seismic images contain valuable subsurface information, yet they are invariably contaminated by random noise that degrades interpretation quality. Existing deep learning approaches for seismic denoising primarily operate on 2D sections or employ attention mechanisms with spatial factorization, windowing, or dimensional reduction that compromise the inherent 3D spatial relationships in seismic volumes. This study introduces a novel architecture combining 3D convolutional neural networks with true global 3D attention, where every voxel attends to all other voxels in the volume without spatial constraints or factorization. Our approach implements a U-Net architecture integrated with global attention modules that compute full 3D cross-correlations through multi-head self-attention mechanisms. Unlike existing methods that apply SWIN transformers or flatten 3D data for computational tractability, our architecture preserves true volumetric relationships by enabling unrestricted voxel-to-voxel interactions across the entire seismic volume. The network combines the spatial locality benefits of 3D convolutions with the global context modeling capabilities of unfactorized attention. To manage the O(N²) computational complexity of voxel-to-voxel attention, we employ sub-volume decomposition (64 × 64 × 64 voxels), mixed-precision training, gradient checkpointing, and multi-GPU parallelism, making unfactorized 3D global attention tractable on modern GPU hardware. Training employs a self-supervised strategy using real-world post-stack migrated seismic volumes from multiple geographically distinct surveys without requiring noise-free ground truth data. Multiple 3D training samples are generated by adding varying levels of random noise (5–30% standard deviation) to existing field-acquired seismic data, enabling the network to distinguish spatially coherent geological features from survey-dependent random noise across different acquisition conditions. This multi-survey training paradigm acts as an implicit regularizer that enhances generalization to unseen seismic data while addressing the fundamental limitation that truly noise-free seismic volumes do not exist in practice. Comparative evaluation against state-of-the-art SWIN transformer-based denoising demonstrates the effectiveness of true 3D global attention for preserving geological structures while suppressing random noise. Robustness is verified through testing on completely unseen surveys from different geological basins, with comprehensive residual analysis confirming preservation of geological content and broadband noise suppression without signal attenuation. Results indicate improved noise reduction performance while preserving subtle geological features that are critical for seismic interpretation. The contribution represents the first application of unfactorized 3D global attention to post-stack seismic denoising, establishing a new paradigm for leveraging complete volumetric context in seismic data processing without reliance on synthetic training data or noise-free ground truth.</p>