Bringing Trust to Edge AI: Our Progress with the REBECCA RISC-V Platform

Artificial intelligence is no longer confined to massive data centers. It is moving out into the field directly into machines, autonomous vehicles, smart infrastructure, and everyday appliances. This shift toward Edge AI cuts down on communication latency, minimizes the volume of data sent to the cloud, and keeps intelligent systems running smoothly even when internet connectivity drops. But bringing AI to the edge creates a difficult balancing act. Edge devices have strict power budgets and limited computing resources, yet the applications demand lightning-fast, high-performance processing. More importantly, when AI is integrated into critical systems, raw performance simply isn't enough. Safety, security, real-time behavior, and strict system isolation become non-negotiable. This is the exact challenge we are tackling in the European REBECCA project.

Building an Open Platform for Edge AI

REBECCA aims to deliver a complete hardware and software stack tailored for efficient Edge AI. Instead of locking into proprietary designs, the project centers around the open RISC-V instruction set architecture. At the core of this hardware is the CVA6, a 64-bit open-source RISC-V processor capable of running full operating systems. REBECCA combines this general-purpose processing capability with specialized accelerators for machine learning. Because the processors and accelerators share an address space, the system can access memory efficiently without moving data around unnecessarily a huge advantage when trying to save power.

Our latest Milestone: PikeOS on a Quad-Core CVA6

Hardware is only half the battle. You need a software foundation capable of controlling those resources and keeping different workloads safely isolated from one another. SYSGO is building the software bridge that connects the project’s hardware with its end applications.

We recently reached a major technical milestone: our team successfully ported the PikeOS real-time operating system and hypervisor to the REBECCA chip, specifically targeting the quad-core CVA6 configuration.

From Code to Hardware: Build & Deployment Workflow

To bring the platform to life on the hardware, we established a complete end-to-end build and deployment pipeline using SYSGO’s CODEO development environment and the REBECCA SDK:

1. PikeOS PSP Build: We configured and built the dedicated Platform Support Package (PSP) tailored specifically to the quad-core CVA6 hardware layout and memory mapping.
2. Partition Projects: We created the individual payload projects—an ELinOS embedded Linux system for rich software services and a lightweight PikeOS native application project for time-critical logic.
3. Kernel Fusion in CODEO: Using CODEO, we combined the hypervisor binary, guest partition images, memory policies, and system layout configurations into a unified Kernel Fusion project.
4. Target Deployment: CODEO compiled the setup into a single bootable binary image. Using the REBECCA SDK, we flashed this final image directly into the board’s QSPI flash memory on the AMD/Xilinx Alveo U55C FPGA, enabling the system to boot autonomously on target startup.

To demonstrate the platform's virtualization and isolation capabilities, we are currently running a dual-partition setup directly on the chip:

• Partition 1: Runs a PikeOS native application, ensuring strict, deterministic real-time execution for critical functions.
• Partition 2: Runs ELinOS, handling standard embedded Linux workloads and broader system tasks.

This hypervisor-driven separation is the key to safe Edge AI. It means a computationally heavy, AI-based image processing application can run on the same physical chip alongside critical control software, without either application sharing an execution environment or stepping on the other's toes.

From the Lab to the real World

Why does this strict separation matter? Because real-world applications demand it. Take autonomous infrastructure inspection as an example. Drones checking bridges or power lines can use local AI to identify cracks or wear instantly. They don't need to beam gigabytes of high-resolution sensor data back to a remote cloud server to make a safe navigational or maintenance decision. The same principles apply to AI-based inspection of photovoltaic panels, high-speed visual checks in semiconductor manufacturing, and intelligent home appliances. In all of these scenarios, data is processed right where it is created. This enables faster decisions, reduces network strain, and removes the dependency on a permanent cloud connection.

The Road ahead

The REBECCA project proves that the future of Edge AI isn’t just about building faster accelerators. It is about creating a cohesive, trustworthy ecosystem where processors, memory, operating systems, and AI frameworks work seamlessly together. By bringing our virtualization, safety, and security expertise to the open RISC-V architecture, we are helping push the industry forward. We aren't just figuring out how to run algorithms at the edge; we are building the foundation for powerful, predictable, and genuinely secure intelligent systems.

More information at www.rebecca-chip.eu