Artificial Intelligence (AI)

AI-supported optimization in SDV: Data → Models → Orchestration → Bench validation.

AI-Powered Optimization in the Software Defined Vehicle

The Software Defined Vehicle Platform integrates modern AI methods to intelligently optimize vehicles, networks, and software functions. Real-time measurement data, bench topologies, and shadow-mode executions provide the foundation for machine learning and graph-based models.

AI thus becomes a central accelerator for software quality, performance, and energy efficiency in the SDV.

AI Pipeline: From Raw Data to Validated Release

The AI pipeline processes real measurement values in four consecutive stages – from raw data to hardware-validated release:

AcquisitionBench · Shadow ModeTrainingGNN · Latency ModelsOptimizationRouting · PlacementValidationBench Validation

01 – Realistic Measurements from Bench & Shadow Mode

The SDVA Test Bench collects high-resolution data at microsecond level:

  • Latency & Jitter: End-to-end measurement per hop, transmission time variation analysis
  • Bus Load: Data rate per bus – Ethernet/TSN utilization, CAN/CAN-FD bus load, error frames
  • Hardware Metrics: CPU/memory utilization, temperature profiles, and energy consumption per node
  • Distributed Tracing: Wasm module interaction with FPGA-based timestamps – µs-precise, versioned
Real Data Instead of Simulation: The AI training pipeline is based on physical measurements from the bench – not synthetic simulation data. This allows the models to learn realistic jitter characteristics, hardware-dependent latencies, and actual system behavior under load.

From Bench to Pipeline

High-resolution measurements from the SDVA Test Bench – latencies, bus loads, temperature profiles – flow directly into the AI training pipeline. No synthetic data, no simplifications.

"What the bench measures, the model learns – physically grounded, not simulated."

Technical illustration: A test bench rack on the left emits glowing data streams in blue and cyan flowing right toward a stylized data store. Measurement icons – waveform, clock, thermometer – float along the data flows, symbolizing the types of metrics being captured.
Bench Data as AI Training Foundation

02 – Graph Neural Networks for SDV Topologies

The bench exports topology graphs (ECUs = nodes, bus connections = edges) and measurement-based latency profiles to train Graph Neural Networks.

The models learn:

  • Latency Prediction: Forecasting end-to-end latencies for new E/E architectures
  • Bottleneck Detection: Identification of critical paths, bottlenecks, and QoS issues
  • Software Placement: Automatic assignment of modules to ECU, core, or accelerator
  • Architecture Evolution: Evaluation of migration paths – domain → zone → central
  • GNN
  • Topology Graphs
  • Latency Prediction
  • Software Placement
  • TSN

Graphs That Understand Vehicles

Graph Neural Networks operate directly on the vehicle's E/E topology: ECUs as nodes, bus connections as edges, latency profiles as weights. This produces models that evaluate architectures before hardware is installed.

"The network understands the network – GNNs detect patterns that remain invisible manually."

Futuristic network diagram on a dark navy background: five large glowing nodes connected by luminous cyan-blue lines. Three central nodes form a triangle; two outer nodes connect from upper-left and lower-right. The nodes and lines emit a bloom glow, visualizing a Graph Neural Network analyzing vehicle topology data.
Graph-Based Architecture Analysis

03 – AI-Based Optimization of Vehicle Architecture

The AI models deliver concrete optimization proposals that are implemented in the SDV Platform:

  • Network Paths & Routing: Minimizing latency and jitter while distributing load across parallel paths
  • Software Placement: Automatic selection of the optimal hardware platform – x86, ARM, RISC-V, GPU, FPGA – based on latency and energy profiles
  • TSN Profiles: Optimization of priorities, VLAN settings, and TAS scheduling for different traffic classes
  • Failover Strategies: Detection of critical nodes and proactive switchover to redundant paths – before the fault occurs
Intelligent Self-Optimization: The combination of GNN predictions and bench validation enables SDV architectures that autonomously adapt to changing requirements – from network topology to function distribution on hardware.

Intelligent Routing in Real Time

AI models analyze the vehicle topology and identify optimal data paths considering latency, utilization, and redundancy. Suboptimal routes are automatically detected and replaced with better alternatives.

"The AI doesn't optimize blindly – it validates every recommendation on the physical bench."

Schematic illustration of a vehicle E/E network topology as a graph on a dark navy background. An AI highlights the optimal data path in bright cyan while suboptimal routes appear dimmed. Small latency indicators are visible at key nodes, quantifying the optimization effect.
AI-Powered Network Routing

04 – Realistic Assurance Under Hardware Conditions

All AI proposals undergo full hardware validation on the SDVA Test Bench before release:

  • Physical Switching: Topology changes via the switching matrix – real signal paths, not virtually routed
  • End-to-End Latency: Measurement of actual delay under real conditions
  • Shadow Mode Comparison: AI proposals run in parallel with production logic – risk-free assurance
  • FPGA Timestamps: Hardware-based time measurement eliminates software overhead – µs precision
Validation before Deployment: Only after successful bench validation is a network layout, a TSN configuration, or a software placement recommended. No AI proposal goes into production without physical assurance.

Why AI Is Becoming Indispensable in the SDV

  • Data-Driven Instead of Static: Modern E/E topologies with hundreds of nodes and thousands of parameters are too complex for manual optimization – AI recognizes patterns that humans overlook
  • Faster Development: Reproducible bench tests and shadow mode validation accelerate the path from prototype to production readiness
  • Performance & Efficiency: Smart routing, intelligent software placement, and automatic hardware selection reduce latency and energy consumption
  • Prerequisite for OTA: Dynamic homologation requires automated risk assessment, dependency analysis, and performance prediction – exactly what the AI pipeline delivers

In Brief

AI makes the SDV faster, safer, and more efficient – from topology to function logic. The platform combines:

  • Real-World Data: High-resolution bench measurements instead of synthetic simulations
  • Learning Models: GNNs that predict latencies and evaluate architectures
  • Automated Decisions: Routing, placement, and TSN configuration driven by AI
  • Hardware Validation: Every AI proposal is physically verified on the bench

The result is an intelligent SDV platform that continuously improves.

More Technology Pages: SDV Platform · WebAssembly · Test Platform