Xilinx Unveils Kria SOMs to Accelerate Edge AI and Embedded Vision

With the launch of its first System‑on‑Module (SOM), Xilinx is redefining how developers bring edge AI and embedded vision to market. The Kria K26 SOM, coupled with the low‑cost KV260 AI Vision Starter Kit, delivers a complete hardware and software solution that eliminates the need for FPGA design expertise.

Designed for smart cities, factories, and security systems, the starter kit is priced at $199, while the commercial and industrial variants of the Kria K26 SOM retail for $250 and $350 respectively. Xilinx’s pricing strategy targets mass‑market adoption and positions the SOM as a turnkey solution for vision AI projects.

Modern vision AI is notoriously complex, and many developers lack deep hardware knowledge. Xilinx addresses this by offering pre‑built hardware platforms paired with industry‑standard software stacks and a growing library of accelerated applications. This abstraction enables millions of developers to integrate AI without chip‑level design work.

The Kria ecosystem provides a self‑enabled journey from exploration to production. Tutorials, training courses, and an ecosystem of providers accelerate the design cycle for hobbyists, makers, and commercial teams alike.

Underpinning this acceleration is Xilinx’s investment in toolflows that make adaptive computing accessible to software developers. The Vitis unified software development platform supports TensorFlow, PyTorch, and Caffe, as well as C, C++, OpenCL, and Python. Developers can plug their custom AI models into pre‑built vision pipelines and deploy them with minimal effort.

Expanding developer familiarity, Xilinx now supports Yocto‑based PetaLinux and, in partnership with Canonical, will provide Ubuntu Linux support. "For smart vision applications, developers and innovators want the Ubuntu experience they’re used to from cloud to desktop," says Thibaut Rouffineau, Vice President of Marketing, Canonical/Ubuntu. "Together with Xilinx, we’re excited to give Kria customers out‑of‑the‑box productivity and a frictionless path to production with guaranteed stability and security."

Save up to nine months of development time

Ready‑to‑deploy modules and applications are part of a broader industry trend to demystify edge AI. According to Jeff Bier, founder of the Edge AI and Vision Alliance, many companies lack dedicated machine‑learning teams. Vendors like Xilinx are filling this gap with reference designs and software libraries.

During the launch event, Chetan Khona, Director of Industrial, Vision and Healthcare, emphasized the speed advantage: "Customers can save up to nine months in development by using a module‑based design. The starter kit lets users start within an hour, with no FPGA experience needed. Connect the camera, insert the pre‑programmed microSD card, power up, and select an accelerated application to run immediately."

Senior executive Kirk Saban highlighted the strategic shift: "Xilinx’s entry into the SOM market extends our reach beyond data‑center chips to embedded systems, making adaptable hardware accessible to millions of software and AI developers."

The Kria K26 SOM is built on the Zynq UltraScale+ MPSoC architecture, featuring a quad‑core Arm Cortex‑A53 processor, over 250,000 logic cells, a H.264/265 codec, 4GB DDR4 memory, and 245 I/O pins. With 1.4 tera‑ops of AI compute, it delivers three‑fold higher performance at lower latency and power than GPU‑based SOMs—ideal for security cameras, traffic monitoring, retail analytics, and vision‑guided robotics.

Complementing the hardware, Xilinx launched the first embedded app store for edge applications. The store offers a library of accelerated apps from Xilinx and ecosystem partners, ranging from smart camera tracking to natural language processing. Xilinx’s open‑source apps are available at no charge.

Early adopters demonstrate the SOM’s impact: Kutleng Engineering Technologies deployed wildlife‑safety cameras in South Africa within two months, while Optimized Solutions Limited in India used the Kria SOM for multi‑object detection in smart‑city applications.