Embedded Generative AI for Software-Defined Vehicles

Embedded generative AI is being integrated into software-defined vehicles through a collaboration between FEV, Microsoft and NVIDIA. The project combines embedded small language models, GPU-accelerated computing and multimodal processing to support voice, text and gesture interactions directly within the vehicle for automotive applications.

Technical Cooperation for Embedded AI
The cooperation brings together complementary expertise in automotive engineering, artificial intelligence and high-performance computing. FEV is responsible for system architecture, AI integration, software development, validation and vehicle deployment. Microsoft provides the Phi-4-mini-instruct small language model through Microsoft Foundry, while NVIDIA supplies the DRIVE AGX computing platform together with AI model optimization tools.

The collaboration addresses a key challenge in software-defined vehicles: delivering reliable AI functions without depending on continuous cloud connectivity. By executing inference locally inside the vehicle, essential services remain available when internet access is unavailable or limited while reducing communication with cloud infrastructure.

Embedded AI Architecture and Responsibilities
The technical solution is based on embedded small language models operating as modular software services on NVIDIA DRIVE AGX hardware. The multimodal architecture processes speech, text and visual inputs to enable natural human-machine interaction.

To improve model efficiency for embedded deployment, FEV fine-tuned Microsoft's Phi-4-mini-instruct model using synthetically generated datasets curated with NVIDIA NeMo. The optimized model is then integrated into the vehicle platform, where GPU acceleration supports low-latency inference within automotive hardware constraints.

The embedded language model complements cloud-based large language models by providing local backup intelligence. This architecture allows vehicle manufacturers to allocate AI workloads between local and cloud resources according to performance, connectivity and operational requirements.

Deployment in Software-Defined Vehicles
The platform has been implemented in demonstration vehicles to validate real-time operation under representative automotive conditions. A dashboard configuration demonstration enables users to personalize vehicle settings through natural voice commands, including dashboard layouts and driver profiles, without requiring permanent internet connectivity.

The embedded platform is designed for integration with existing vehicle electronic architectures and supports gradual deployment of locally executed AI functions alongside cloud-based services. Validation results will support future customer projects and near-series vehicle applications.

Industrial Applications
The cooperation targets multiple automotive applications, including SAE Level 3 to Level 5 automated driving, driver and occupant monitoring, and personalized human-machine interface configuration.

Embedded generative AI can improve object recognition, traffic scene interpretation and route analysis in complex driving environments. Local processing also supports fatigue and distraction detection while maintaining availability of safety-related functions during connectivity interruptions. Voice-controlled vehicle personalization enables rapid configuration of user profiles and interface settings without relying on external servers.

Expected Technical Impact
Running AI inference directly inside the vehicle improves response time and operational resilience while reducing dependence on cloud infrastructure. The combination of embedded small language models with cloud-based AI provides a scalable architecture that balances computational performance, connectivity requirements and infrastructure costs for future digital mobility platforms.

Edited by Evgeny Churilov, Induportals Media - Adapted by AI.

www.fev.com