Mandali Khalesi
My journey into vision AI and cloud engineering began over a decade ago during my time managing APAC automotive products at HERE Technologies. Back then, “intelligence” largely referred to business analytics and cloud adoption was primarily focused on delivering data services efficiently to the edge. I saw firsthand how predictive analytics— powered by driver location, time of day and real-time traffic data—could help civil engineers in Melbourne redesign roads or enable commuters in Bangkok to avoid congestion before it formed. These early experiences reinforced my belief in the power of timely, data-driven insights delivered to edge devices.
My passion for AI deepened during my tenure as General Manager of Automated Driving Mobility and Innovation at Toyota Motor Corporation and later as Vice President of Automated Driving at Woven Planet. I witnessed the evolution of ADAS—from mono-camera systems to high-resolution, sensor-fused platforms capable of true situational awareness. We used cloud-based machine learning to analyze vast databases of raw satellite imagery, constructing detailed road networks that enabled self-driving cars to navigate unfamiliar environments safely. While we initially relied on rule-based programming, the transition to machine learning for perception and path planning marked a significant leap, even at a time when AI was still in its infancy.
The emergence of transformer-based models like ChatGPT marked another major inflection point. The automotive industry began moving beyond traditional CNNs and rudimentary voice assistants toward transformer architectures for both infotainment and safety applications. As Elon Musk recently commented to Yann LeCun, CNNs are becoming increasingly obsolete in the face of this shift. Today, the trend is clear: we’re moving from “bigger is better” to “leaner is better.” The rise of small language models like TinyLlama, Phi-2 and GPT-4o-mini is a welcome development for those of us working at the edge—where power, data, compute and cost constraints are ever-present. Yet, realizing the full potential of AI in these environments remains a work in progress.
One of the persistent challenges in deploying new AI models is hardware flexibility. Architectures that can adapt to changing models have consistently outlasted those built around rigid assumptions. The key lies in developing scalable microarchitectures composed of simple, reusable compute nodes—scalar, vector and tensor elements. This would liberate hardware teams from having to predict the future four years in advance, allowing them instead to assemble hardware by instantiating software models from a shared pool of elements. The result: scalable compute resources, consistency across product lines and faster time-to-market.
“The interplay between edge processing and cloud intelligence is redefining the future of AI inference, particularly in computer vision”
Thermal constraints in vision SoCs have also posed significant challenges to energy-efficient AI. At Renesas, we’ve leaned heavily on open-source solutions. We built our HyCo AI compiler on Apache TVM, allowing us to customize AI models without being locked into proprietary vendor tools. This compiler now supports workloads for our fourth-generation SoCs. In collaboration with Japan’s NEDO agency, we’ve demonstrated the power of hardware-software co-design through Hardware Neural Architecture Search (HW-NAS), which iteratively determines the optimal hardware and software configuration for a given application. This approach has further reduced power consumption—critical for edge AI solutions in automotive and industrial settings.
In parallel, open-source hardware is playing a growing role in AI innovation. We recently introduced our first RISC-V MCU products and continue to actively engage with the global RISC-V and open hardware community. I regularly participate in grassroots RISC-V conferences in Japan and firmly believe that an expanded open-source hardware ecosystem will accelerate innovation, much like open-source software has over the past 30 years.
The interplay between edge processing and cloud intelligence is redefining the future of AI inference, particularly in computer vision. At the lower end of the spectrum, microcontrollers (MCUs) are valued for their simplicity and power efficiency, but they struggle with the increasingly complex workloads demanded by modern AI models. As transformer-based architectures become the norm, MCUs must begin to behave more like MPUs. Our legacy in lean automotive MCUs gives us a strong foundation for this transition. At higher compute levels, our MPUs and third-generation SoCs and beyond are already capable of supporting ADAS with real-time predictive capabilities.
The key to scaling lies in unifying the AI software stack across MCUs, MPUs and SoCs. This enables code reuse, reduces development time and accelerates innovation. Our cloud-based AI Workbench software, announced at CES 2024, is a step in that direction. It allows developers to create and optimize AI applications across multiple hardware platforms via a simple browser interface— streamlining the path from concept to edge deployment.
According to MarketsandMarkets, the edge AI inference market is expected to grow from $100 billion in 2025 to $250 billion by 2030—far outpacing the growth of AI training. A consistent hardware and software architecture that bridges automotive and industrial use cases will be critical in addressing this multi-billiondollar opportunity.
Scaling AI inference while maintaining ultra-low power consumption requires a careful balance of performance and efficiency. Most models are trained in data centers with abundant resources, resulting in bloated architectures ill-suited for edge deployment. That’s why we’re integrating our in-house MLOps tools and quantization techniques into the AI Workbench. By providing developers with both performance metrics and visual comparisons of edge cases, we help ensure their models perform reliably in real-world scenarios.
To aspiring AI and cloud engineers, my advice is twofold. First, commit to continuous learning and unlearning. At Renesas, we launched “Coder Fridays,” a structured initiative that gives engineers dedicated time for upskilling, supported by a customized curriculum and nanodegree programs. This investment in growth drives both innovation and technical excellence. Second, stay engaged with open-source communities. Whether in software or hardware, collaborative innovation is accelerating at an unprecedented pace. Staying connected to these ecosystems is essential for anyone looking to stay at the forefront of technology.
