(By Guilherme Marshall, Director of Automotive Business for Arm in Europe, the Middle East, and Africa) Close your eyes and imagine you are getting into a car to go to work. The temperature and cabin settings have already been adjusted to your preferences. Once inside, the car, having learned your driving habits, proactively asks if you are heading to the office again. Although your car currently cannot navigate unsupervised in the busy urban area where you live, you can take your hands off the steering wheel and let the car drive autonomously in congested traffic, while you simply pay attention to the road conditions.
Before long, your car will inform you: you can relax now, there’s no need to keep your eyes on the road; it will take you to the office in about 45 minutes. During this time, you can browse the news, handle emails, attend two meetings, or enjoy the last episode of the series you didn’t finish last night.
These scenarios will be realized through in-car artificial intelligence (AI) technology. The previous predictions about fully autonomous vehicles navigating city streets have given way to a more pragmatic development path. Automakers are currently focusing on expanding the operational design domain (ODD) of Level 2+/++ driver control assistance systems (DCAS) and Level 3 automated lane-keeping systems (ALKS) to pave the way for higher levels of autonomy.
However, the large-scale deployment of autonomous driving still faces numerous challenges, including:
• Algorithm limitations
• High development costs and long development cycles
• Hardware constraints
Accelerating Autonomous Driving with Arm Zena Computing Subsystem
Despite these limitations, there is a feasible path to achieving Level 2++ and Level 3 driving capabilities—this path is being developed collaboratively by the Arm Zena Computing Subsystem (CSS) and partners across the automotive ecosystem. Currently, Zena CSS has been adopted by leading automakers and is set to be integrated into cutting-edge autonomous driving control units, soon to be deployed in public transportation. In fact, leveraging Zena CSS, industry leaders believe they can exceed analyst predictions and significantly accelerate the mass production process from Level 2+ to Level 4 vehicles.
Breaking Algorithm Limitations: Tailoring Technology for the AI-Defined Automotive Era
Operating autonomous vehicles on the road is a multidimensional and complex task. The industry has invested billions of dollars to translate clear safety standards into executable program logic, which defines how autonomous vehicles should operate under known safe states and known risks within the ODD. However, as the ODD expands and new markets emerge, the “long tail effect” of numerous known and unknown risks has proven to be too costly and difficult to achieve safe, scalable applications.
While most risks are predictable, there will still be some unknown risks that make it difficult for vehicles to respond appropriately. For example, a car can react to a cow appearing on the road and adjust its driving state accordingly, but how should it respond to a giant cow model being transported on a trailer? Such questions are critical considerations when designing autonomous driving systems.
Fortunately, recent breakthroughs in AI provide automakers with new tools to build sustainable technology models, enabling them to launch more advanced autonomous driving systems in more markets faster. With attention-based neural networks and visual-language-action models (VLAM), end-to-end learning technologies open a technical path for autonomous driving systems to better handle edge cases, making decisions through contextual reasoning even in previously unseen scenarios. Meanwhile, self-supervised learning significantly reduces the time and costs currently spent on data organization and labeling. With appropriate safety measures in place, these new technologies will help create autonomous driving systems with generalization capabilities that truly meet the scaling demands of global automakers.
To learn more about end-to-end AI technology in the autonomous driving field, listen to the conversation between Suraj Gajendra, Vice President of Products and Solutions at Arm Automotive, and Silvius Rus, Vice President of Software at Wayve.
Zena CSS is meticulously designed for the next generation of autonomous driving systems defined by AI in vehicles. This thoroughly validated computing subsystem features seamless system scalability, from integration into entry-level in-vehicle infotainment (IVI) systems and advanced driver assistance systems (ADAS) to L2+, L3, and L4 domain control units. Zena CSS is built on Neoverse CMN S3AE automotive mesh network interconnect technology, natively supporting chiplet architecture, and can easily integrate heterogeneous computing chiplets such as GPUs and AI accelerators through standardized UCIe interfaces, facilitating the creation of customized, adaptable products.
Zena CSS is equipped with a 16-core Arm Cortex-A720AE CPU, integrating powerful AI computing capabilities within its high-performance processing module. These CPUs, based on the Armv9 architecture, introduce new instructions that significantly accelerate workloads such as AI and computer vision (CV). Developers can seamlessly access related functionalities through the open-source Arm KleidiAI software library. For instance, in perception benchmarks (such as point cloud conversion and bird’s-eye view construction), when the vehicle’s workload is migrated from the Arm Cortex-A78AE CPU to the Cortex-A720AE, performance improves by 30%.
Additionally, by optionally integrating the innovative Arm Mali-C720AE ISP, AI perception performance can be further enhanced. Traditional ISPs often require lengthy and semi-manual processes to optimize for human visual effects; however, the differentiable ISP functionality model of Mali-C720AE can conduct closed-loop training with camera perception models, automatically generating ideal configurations for users’ specific computer vision technology stacks. Furthermore, the dual pipeline support of Mali-C720AE allows for parameter configuration for both human visual pipelines and computer vision pipelines, thus providing better performance for both usage scenarios.
Solving Cost and Cycle Challenges: Reducing Development Time by Up to 12 Months
The next-generation system-on-chip (SoC) for autonomous driving is a highly complex semiconductor device. Both hardware and software development require significant investment, and even well-funded global companies feel the pressure. The limited reusability of hardware and software across different vehicle platforms exacerbates this issue—not only does it lead to exponentially increased costs, but it also extends the time to market. Zena CSS directly addresses these challenges by standardizing non-differentiated building blocks, optimizing ecosystem investments, and simplifying porting processes, effectively resolving the aforementioned issues.
As Arm’s Senior Vice President and General Manager of Automotive, Dipti Vachani, stated, in the AI era, automakers must maintain competitive advantage by ensuring safety, energy efficiency, and flexibility while also possessing scalable computing capabilities. Zena CSS can shorten chip development time by up to 12 months and reduce chip engineering investments for each project by up to 20%, helping automakers and chip suppliers bring new vehicles to market faster.
Zena CSS integrates validated low-power high-performance CPUs, dedicated security islands, runtime information security enclaves, reference firmware, and software support, forming a complete subsystem ready for chip implementation. Importantly, it also provides relevant certifications compliant with ISO 26262 (functional safety) and ISO 21434 (cybersecurity).
To accelerate software development, Arm collaborates with ecosystem partners to provide subsystem-level virtual platforms for Zena CSS even before hardware readiness. Overall, the standardization of the computing subsystem and foundational software layer can reduce cross-platform porting workloads by up to 30%.
Breaking Hardware Limitations: Redefining Energy Efficiency
Today, automakers have found a viable path: designing a computing architecture that can scale across vehicle classes and ODD while meeting stringent cost and energy efficiency requirements. Zena CSS adopts a modular and scalable design philosophy, leveraging Arm’s leading advantages in energy efficiency and mature licensing business models. By reducing the homogenization costs within the industry, SoCs built on Zena CSS will have a competitive edge, allowing automakers to focus their R&D investments more precisely on core capabilities that create higher value.
Advancing Towards Autonomous Driving
Whether supporting Level 2+ DCAS for compact vehicles or enabling Level 4 highway autonomous driving capabilities for high-end models, Zena CSS provides automakers with an ideal path to create a unified autonomous driving computing platform that covers all vehicle models. This platform not only reduces integration workloads and shortens time-to-market but also lowers validation and compliance certification costs, thereby accelerating the realization of the vision for scalable autonomous driving.
With Zena CSS at its core, the future development prospects of autonomous driving are promising. Arm looks forward to witnessing the autonomous driving computing systems tailored for AI-defined vehicles shine brightly in the process of accelerating the large-scale deployment of autonomous driving technology.