This paper analyzes the I4U audio systems developed for NIST SRE24. We implemented and evaluated state of the art DNN embedding models, including established architectures and a new Multi2dCoTNet topology, which delivered reliable performance gains across all conditions. We used advanced embedding classifiers, score normalization, and an original Mixed Norm data augmentation technique. In this paper we outline the contribution of each component. In addition, building upon recent literature, we developed an accurate system that leverages embeddings extracted by a self supervised model and achieves competitive accuracy. When fused with our best Multi2dCoTNet embedding extractor, it yields a 2.5% relative reduction in minimum primary cost compared to our full open training submission, with 69% computational cost reduction. We further assess this simpler and deployable two model solution across gender, source type, language match, and duration conditions in SRE24.