Access Official Technical PDF

Share Summary:LinkedInEmail

Automotive computing platforms are being reshaped by functions such as automated driving, vehicle connectivity and online optimization of electrified powertrains. These applications bring greater computational demand and connect software that may have very different timing and safety requirements. As vehicle architectures consolidate functions onto shared hardware, engineers must consider not only how much work a platform can perform, but whether it can do so predictably and safely.

That challenge was central to discussions at the 2019 Dagstuhl Seminar on Future Automotive HW/SW Platform Design. Its themes remain relevant to system architects and integrators working with complex embedded systems: establish predictable behavior, integrate mixed-criticality components safely, and make heterogeneous hardware programmable without losing control of system-level properties.

Why automotive computing platforms are changing

Traditional vehicle architectures distribute functions across electronic control units (ECUs), often with software and hardware configured for specific roles. New functionality and rising software complexity are encouraging a shift toward centralized or domain-based computing, where applications that were previously separated may share a more capable platform.

These platforms can combine general-purpose application cores, safety-oriented cores, GPUs and other accelerators. They may also connect to in-vehicle networks and, in connected applications, to external infrastructure. The result is a mix of computing resources, communication paths and application models. Control, stream processing and perception or cognition workloads, for example, do not necessarily share the same execution or timing needs.

Consolidation can change how teams allocate resources, manage software updates and define system boundaries. But it also means that the behavior of one application may depend on shared processors, memory, communication links or operating-system services. A platform’s headline processing capability does not, by itself, establish that time-critical functions will meet their deadlines or remain isolated from other workloads.

Predictability matters as much as processing performance

Many high-performance components were not originally designed for safety-critical, hard real-time use. Their computational capabilities may be well understood, while details that affect timing—such as scheduling behavior or contention for shared resources—can be harder to characterize. Engineers therefore need to assess extra-functional properties, including determinism, predictable response times and freedom from interference.

On a heterogeneous system-on-chip, execution time can depend on how work is scheduled across processors and accelerators, and on competition for caches or system memory. If those effects are not understood, it becomes difficult to perform reliable timing analysis or make defensible guarantees for critical tasks. The design question is not simply which accelerator is fastest, but whether its behavior can be modeled and controlled within the requirements of the application.

Possible approaches discussed in the seminar included making resource access more predictable, analyzing accelerator scheduling, and using task models that represent parallel work and timing constraints. Weakly-hard timing constraints offer another way to describe systems in which a bounded number of deadline misses may be acceptable. Such models can support analysis of adaptation and resource use, but they must be applied with care: the permitted timing behavior has to remain consistent with the function’s safety requirements.

Safe integration requires clear models and boundaries

When applications with different criticality levels share a platform, architecture teams must address how they interact and how failures or resource demands are contained. Relevant design choices include operating-system configuration, task scheduling, virtualization, communication middleware and monitoring. Safety and security also need to be considered together, particularly as vehicle software becomes more connected and includes learning or adaptive components.

Centralizing functions may reduce some forms of distributed coordination, but it does not remove the need for system-level safety engineering. A change to one application or requirement can affect other software, hardware allocation and safety evidence. As vehicle variants and software configurations multiply, teams need traceability that helps identify which analyses and assurance artifacts are affected by a change.

Communication is part of the same integration problem. Service-oriented and data-centric architectures can support applications that exchange information across distributed components, but in-vehicle communication must still meet requirements for latency, prioritization, safety and security. Time synchronization and coordinated scheduling are among the concepts explored for establishing system-wide timing behavior.

Programming models connect applications to hardware

Heterogeneous hardware is useful only if software can exploit it without making system behavior opaque. Programming abstractions and models of computation can describe how an application is structured, how it handles time and how it maps to available resources. Dataflow models, for instance, can provide structure for parallel execution; reactive and timed models can help express event-driven behavior and timing relationships.

These abstractions matter beyond developer convenience. A well-defined model can support analysis, mapping and runtime decisions about resource use. However, the model must represent the application’s real behavior, and information needs to flow across the software stack—from the application and middleware to the operating system and hardware. A mismatch between programming assumptions and execution behavior can undermine predictability.

Shared terminology is also important. The seminar highlighted that terms such as “execution model” can be used differently across technical communities. For engineering teams spanning software, hardware, controls and safety, agreeing on what each model describes helps reduce ambiguity during architecture design and integration.

Validation depends on representative workloads

Simulation and benchmarking can help teams evaluate design choices before deployment. Co-simulation approaches can combine models of software and physical components, while benchmarks can provide workloads for comparing programming models and heterogeneous platforms. Their value depends on how well they represent relevant automotive applications and constraints.

A benchmark that measures processing speed alone may not answer questions about timing predictability, maintainability or the effort needed to program a platform. The seminar discussions emphasized the importance of industry-relevant benchmarks so that research and engineering approaches can be assessed against practical needs. For integrators, evaluation should include non-functional behavior and system-level interactions—not just isolated component performance.

Designing for dependable performance

For engineering teams, the practical task is to treat platform selection, software architecture and verification as connected decisions. Define timing and safety requirements early; identify shared resources and possible interference; select programming and communication models that make behavior analyzable; and validate the integrated system with workloads that reflect its intended use.

Automotive computing platforms can support increasingly demanding functions, but capability alone is not a sufficient design criterion. Dependable deployment depends on understanding how applications execute, communicate and share resources—and on maintaining evidence that the integrated system meets its requirements. By bringing predictability, safe integration and validation into architecture decisions from the outset, teams can make more informed trade-offs as vehicle hardware and software continue to converge.

Access Official Technical PDF

About The Author