Achieving ASIL Compliance in Automotive AI Systems

By Arteris

June 11, 2026

Key Takeaways

• There is currently no single industry-standard approach to achieving ASIL-compliant AI in automotive systems.
• Most automotive architectures today keep AI accelerators outside the highest safety boundary and rely on ASIL D safety islands to validate outputs.
• AI accelerator vendors are increasingly adding ASIL A, B, or C capabilities directly into their products, but approaches vary widely.
• Redundancyremains a foundational safety strategy, whether through duplicate processors, multiple sensors, or independent validation paths.
• Regardless of the approach, safe and deterministic communication between AI, compute, memory, and safety domains is essential.

The Challenge of Making AI Safe Enough for Cars

Artificial intelligence is rapidly becoming a foundational technologyin modern vehicles. Advanced driver assistance systems (ADAS), automated parking, driver monitoring, sensor fusion, and autonomous driving all depend on AI models that can process enormous volumes of data from cameras, radar, LiDAR, and other sensors in real time. Delivering this level of performance requires specialized AI accelerators capable of executing complex neural networks efficiently within the power, thermal, and cost constraints of an automotive system.

As AI takes on a larger role in vehicle decision-making, however, a new challenge emerges: functional safety. Automotive systems must meet stringent safety requirements defined by ISO 26262, with many ADAS and autonomous driving functions targeting ASIL D, the highest Automotive Safety Integrity Level. This creates a unique engineering challenge because AI accelerators are designed to perform probabilistic inference, while functional safety frameworks are built around deterministic behavior that can be verified and validated.

The challenge is that AI and functional safety were never designed toworktogether, and AI inference engines are fundamentally probabilistic. They are often described as “black boxes” because understanding exactly how a neural network arrived at a specific conclusion can be difficult, even for the engineers who designed it.

This means that as AI becomes more deeply integrated into vehicle operation, the industry faces a critical question:How do you build an automotive system that meets ASIL D safety goals when part of that system relies on non-deterministic AI?

The answer today is not a single architecture, as the industry continues to explore several different approaches.

The Industry Has Not Yet Converged on a Single Solution

Unlike CPUs, microcontrollers, and other established automotive processing elements, AI accelerators are still relatively new to the functional safety landscape.As a result, there is currently no universally accepted model for how AI should operate in an ASIL-compliant automotive system. Different semiconductor companies, OEMs, and Tier 1 suppliers are pursuing strategies tailored to their performance goals, risk tolerance, system architecture, and safety philosophy.

While some vendors emphasize external validation mechanisms, others focus on redundancy, and some are beginning to build functional safety features directly into their AI accelerators. This lack of consensus suggests the industry is still in an exploratory phase, where multiple architectural approaches are being evaluated in parallel.

The Established Model: Keep AI Outside the Safety Boundary

Today, the most common approach treats the AI accelerator primarily as a high-performance compute engine, rather than a safety-certified decision-making element.

In this architecture, AI processes sensor data and generates outputs, but those outputs are not trusted blindly. Instead, they are passed to an independent safety processor or safety island that performs deterministic validation before any vehicle action is taken.

However, the safety island does not attempt to understand the neural network’s internal reasoning. Instead, it evaluates whether the resulting action is reasonable and safe.For example, if an AI system recommends an aggressive steering maneuver while the vehicle is traveling at highway speed, a safety validator can determine whether that action falls within acceptable operating parameters before allowing it to influence vehicle behavior.

This approach aligns well with traditional functional safety principles because it avoids the challenge of certifying the neural network itself, while still providing a mechanism for detecting potentially unsafe outputs.Many automotive architects continue to view this safety-island model as the most mature and practical solution available today.

Functional Safety Is Gradually Moving Closer to the AI

While the safety-island approach remains common, the market is beginning to evolve, and a growing number of AI accelerator vendors are introducing functional safety capabilities directly into their products. Rather than positioning AI as entirely outside the safety architecture, these vendors are incorporating varying levels of safety functionality within the accelerator itself.

Some solutions target ASIL A, others ASIL B, and some ASIL C. Each approach reflects a different balance between safety coverage, silicon area, power consumption, performance, and implementation cost.

To this point, the economicsare important. Functional safety mechanisms consume silicon area and add complexity, while higher ASIL targets require increasing levels of fault detection, monitoring, and redundancy. Moving all the way to ASIL D within an AI accelerator would be extremely expensive, which may explain why the industry has largely focused on lower ASIL levels within the accelerator itself and why it faces a growing diversity of approaches rather than convergence around a single model.

Redundancy Remains a Powerful Safety Strategy

When dealing with systems that cannot always be analyzed deterministically, redundancy becomes particularly valuable.Instead of relying on a single processing path, redundancy allows systems to compare multiple independent results and identify inconsistencies that may indicate faults.This can take several forms:

• Duplicate AI accelerators processing the same workload
• Independent compute paths comparing results
• Safety processors performing cross-checks
• Diverse neural network implementations
• Multiple sensor modalities validating one another

Sensor fusionis particularly important in automotive systems. While some autonomous driving strategies rely heavily on cameras, many architectures fuse cameras, radar, and LiDAR to provide multiple perspectives of the vehicle’s environment. If one sensor experiences degradation or produces an unexpected result, other sensors can help validate the situation. Rather than attempting to prove the correctness of a single processing engine, redundancy provides a practical way to improve overall system robustness.

As automotive AI systems become increasingly complex, redundancy is likely to remain a central component of functional safety strategies.

Safety Is Ultimately a System-Level Problem

One of the most important observations emerging from the industry is that functional safety cannot be viewed solely through the lens of an individual accelerator.Safety exists at multiple levels simultaneously: An AI accelerator may have its own safety mechanisms. A CPU cluster may have a different safety target. A safety island may operate at ASIL D. Ultimately, however, regulators, OEMs, and consumers care about the safety of the complete vehicle.

This creates a hierarchical safety model where the requirements of each subsystem are determined by its contribution to overall vehicle behavior, and viewed through this lens, the question may not be whether an AI accelerator itself can achieve ASIL D. The more relevant question may be how AI accelerators, safety processors, CPUs, sensors, memory systems, and software work together to achieve ASIL D objectives at the vehicle level.

Explore Arteris IP:

• FlexGen Smart Network-on-Chip (NoC) IP
• FlexNoC Interconnect IP
• CodaCache Last-Level Cache IP
• Ncore Cache Coherent Interconnect IP

Why Data Movement Matters

Although safety discussions often focus on processors and algorithms, every architectural approach discussed above depends on one common requirement: trusted communication.As automotive AI systems become more heterogeneous, maintaining safe, deterministic, and observable communication between these domains becomes increasingly important.

AI accelerators must exchange data with sensors, CPUs, memory subsystems, safety islands, and other processing domains; safety validators must receive data reliably and within deterministic timing constraints; redundant processing paths must remain synchronized, and safety mechanisms must be able to communicate fault conditions throughout the system.

Whether a vehicle relies on external validation, embedded safety capabilities, redundancy, or a combination of all three, the movement of data between these domains ultimately determines how effectively the overall safety architecture functions.

The industry’s search for an ASIL D AI solution may ultimately reveal that safety cannot be achieved solely within the AI accelerator. As automotive architectures grow more complex, functional safety is increasingly a system-level challenge spanning AI engines, compute domains, safety islands, sensors, memory systems, and software. From independent safety validation to embedded ASIL capabilities and redundant processing paths, leading companies are pursuing different strategies, suggestingthe industry is still determining what “safe AI” should look like in practice. 

The eventual winners may not be those with the most powerful AI accelerators, but those that build the most effective architecture for making AI trustworthy at vehicle scale.

Frequently Asked Questions

Can an AI accelerator be ASIL D compliant?

Today, most automotive architectures do not attempt to make the AI accelerator itself ASIL D. Instead, ASIL D objectives are typically achieved through system-level architectures that combine AI processing with safety islands, validation mechanisms, redundancy, and deterministic control systems.

Why is AI difficult to certify for functional safety?

AI inference engines are probabilistic rather than deterministic. Traditional functional safety methodologies were developed around systems whose behavior could be fully analyzed and verified, making AI a fundamentally different challenge.

What is a safety island?

A safety island is an independent safety processor that validates system behavior and performs safety-critical checks. In many automotive AI architectures, AI outputs are evaluated before they can influence vehicle operation.

Why do some AI accelerators claim ASIL A, B, or C?

Vendors are increasingly integrating functional safety mechanisms into AI accelerators. Different ASIL targets represent different levels of safety coverage and implementation complexity, allowing architects to balance safety, performance, area, and cost requirements.

Is redundancy still important in automotive AI?

Yes. Redundancy remains one of the most effective ways to improve system robustness. It can be implemented through duplicate processors, diverse sensors, independent processing chains, or validation mechanisms that cross-check results.