µ-VF: Enabling Virtualization of Embedded FPGAs

What is it about?

This work presents µ-VF, a framework that rethinks how embedded FPGAs are used at the edge. Instead of dedicating an FPGA to a single, fixed application, µ-VF allows multiple applications to coexist on the same device, each with its own isolated view of the hardware. By enabling on-demand access to acceleration and physical I/O directly on the device, µ-VF turns embedded FPGAs into flexible, shareable edge resources.

Featured Image

Photo by Hazel Z on Unsplash

Photo by Hazel Z on Unsplash

Why is it important?

Many edge applications, such as advanced driver-assistance systems (ADAS), autonomous vehicles, robotics, and industrial control, require fast and predictable data processing together with direct interaction with the physical environment. In these scenarios, relying on cloud resources is often impractical due to latency, bandwidth, or reliability constraints. As a result, computation and hardware acceleration must happen close to sensors and actuators, directly at the edge. While embedded FPGAs are well suited for this role, they are still commonly used in a static, single-purpose way or managed through external hosts. This limits flexibility and makes it difficult to support multiple services that evolve over time. µ-VF addresses this limitation through the concept of a virtual FPGA (vFPGA), where each application is given the abstraction of its own FPGA instance, even though the underlying physical device is shared among multiple users or applications. Unlike approaches that focus only on dynamically offloading short-lived acceleration tasks, µ-VF supports persistent virtual FPGA instances. Each vFPGA behaves as a long-lived virtual device that can continuously run, maintain state, and interact with the external world, while sharing the same physical FPGA with other isolated instances. By supporting both computation and virtualized access to physical I/O, µ-VF enables hardware accelerators not only to process data, but also to sense and act on the real environment. Within an on-device Infrastructure-as-a-Service (IaaS) model, these virtual FPGA instances can be allocated, shared, and released on demand, while retaining direct interaction with sensors and actuators. This improves hardware utilization, simplifies deployment, and enables edge platforms capable of supporting safety-critical and evolving applications side by side.