Zenseact, in collaboration with Embedl, is enhancing autonomous driving technologies by creating an advanced vision stack for ADAS and AD systems. This effort addresses the key challenges of managing memory, energy, and computational efficiency in deep learning models. The approach emphasizes there’s no one-size-fits-all solution; rather, success relies on continuous innovation and the expertise of deep learning specialists and engineers. Zenseact adopts multiple innovative strategies to streamline deep learning applications, such as custom C++ and CUDA implementations for faster execution and multitask deep learning models to reduce computational redundancy. These models combine tasks like semantic segmentation and lane detection into a single network, optimizing resource use. Additionally, hardware-aware model optimization, enabled by the Embedl Model Optimization SDK, further refines performance by tailoring model architectures to leverage hardware capabilities fully, resulting in significant improvements in efficiency and execution speed. We explore many of these aspects in our talk, with a special focus on model optimization.