For the First Time, Xilinx Breaks Into the System-on-Module (SOM) Market
When developing hardware systems for artificial intelligence applications, one of the most crucial considerations is time to market. It’s for this reason that many companies choose to use FPGAs over ASICs; FGPAs offer quicker time to market, more flexibility, and the ability to easily iterate on a design.
Yet, FGPAs aren’t a perfect solution either, requiring a mastery of esoteric hardware description languages (HDLs). For those who need adaptability and fast times to market without learning HDLs, another solution is a system-on-a-module (SOM).
Today, Xilinx—trailblazer of the FPGA—announced that it is . All About Circuits sat down with Chetan Khona and Evan Leal of Xilinx to hear more details about the release firsthand.
What Is a SOM?
Not unlike an SoC, which integrates a number of different ICs onto a single chip, —including SoCs, GPUs, FPGAs, memory, power, and peripherals—onto a small, pre-made PCB.
Basic breakdown of the K26 SOM. Image used courtesy of
The benefit of SOMs is clear: engineers no longer need to spend massive amounts of time integrating all of their components onto a PCB and can instead start with a board in hand.
“SOMs essentially abstract away the hardware so developers can design at the board level instead of having to design at the chip level,” explains Evan Leal, Xilinx's director of product marketing.
“Hardware designers tend to like SOMs because they can avoid the lower-value design work, like memory interfaces. Software developers tend to love SOMs because they can start their work really early, either in parallel or even before the hardware in some cases.”
According to Xilinx, the market for SOMs is growing at roughly 11% year over year with a target of about $2.3 billion by 2025.
Xilinx Kria—A New Family of SOMs
The big news out of Xilinx today is the release of Kria, the company's new family of SOMs. The , which has been designed specifically for vision-based AI applications for edge deployment.
From a hardware perspective, the K26 is based on Xilinx’s and includes an Arm Cortex–A53 Quad-Core subsystem. The system supports 4K 60p video codec and can achieve 1.4 TOPS at inference. Some built-in peripherals include 4 GB of 64-bit DDR4 memory, 40 G Ethernet connectivity, and four USB connections. With 245 IOs, the K26 is flexible enough to operate with any interface or sensor.
Block diagram of the K26. Image used courtesy of
For edge AI, power is a major consideration. Xilinx's director of industrial, vision, healthcare, and sciences—and a co-creator of the Kria SOM—Chetan Khona breaks down the K26’s power performance: “Typical applications that will run on the K26 are going to run under 10 watts. At normal steady-state, you’re talking 8 watts or below for most cases. In a very intensive application, you could be looking at as high as about 15 watts in general.”
Vision AI in South Africa and Beyond
When creating Kria SOMs, Xilinx was faced with a question: What does the vision market need to cut through the design complexity of AI applications? Three requirements became clear:
- Pre-built platforms with both hardware and software solutions for faster deployment
- Flexibility to customize end products for different use cases
- Tools to accelerate AI models for lower cost and latency
Depiction of a Kria SOM used for a facial recognition application. Image used courtesy of Xilinx
With these requirements in mind, Xilinx sees the now-completed K26 finding a home in applications including high-speed object detection in smart traffic cameras, retail analytics (including object tracking and identification), and visual inspection in smart factories.
There have already been success stories with the K26. For example, Kutleng Engineering Technologies is using Kria-based smart cameras to track wildlife in South African National parks. The company says it was able to launch its products within two months as a result of using Kria.
Straightforward Development—Even With No FPGA Experience
Xilinx has fortified this first SOM release with a host of development materials and resources. For one, Xilinx has taken an “accelerated application” approach with this new portfolio, meaning software developers can use top design tools like TensorFlow, Pytorch, or Caffe—along with Python, OpenCL, C, or C++—to input custom AI models and application code. Kria SOMs also support Yocto-based PetaLinux and Ubuntu Linux for embedded customization.
The Vitis development environment is said to provide designers more flexibility in AI development. Image used courtesy of Xilinx
Xilinx has teamed up with its ecosystem partners to expand offerings in the Xilinx App Store—the so-called “first embedded app store for edge applications”—for Kria SOMs. These may range from facial detection to natural language processing, and more.
Kria KV260 Vision AI Starter Kit. Image used courtesy of Xilinx
Along with the K26, the company has also released a starter kit for evaluation and development. Xilinx says the is an out-of-the-box platform for designing vision AI applications, allowing designers—even those with no FPGA knowledge—to jump into a project in under an hour.