Xilinx moves from chip level to board level with new SoMs

Xilinx announced its first foray into the system-on-module market with the unveiling of the Kria portfolio of adaptive SoMs, the first of which will be targeted at helping customers speed up the development of vision AI applications in smart factories and smart cities.

The company described the SoMs as “production-ready small form factor embedded boards” that can be coupled with Xilinx’s complete software stack and even pre-built applications through a new Xilinx app store to help enable faster, cheaper deployment of adaptive computing for edge-based applications. 

“As a chip company, people may look at us like everything we do is complex and expensive, but we are out to subvert that notion,” said Evan Leal, director of product marketing, boards & kits, at Xilinx. “With a SoM, developers will be able to start developing at the board level rather than the chip level.”

What that means, according to Chetan Khona, director of industrial, vision, healthcare and sciences at Xilinx, is that “you are propelled through the design phase much faster because you are starting at a more advanced point.” In some cases, developers may cut down development time by as much as nine months. 

“There is also a cost savings in combining different products at the board level,” Khona said.

 The first product available in the company’s Kria SOM portfolio is the K26 SoM, which will come in both commercial and industrial versions supporting vision AI applications such as security monitoring, traffic and city cameras, retail analytics, machine vision, and vision-guided robotics. The commercial flavor of the K26 will start shipping next month, with the industrial-grade version coming later this summer.

Xilinix said the K26 is “built on top of the Zynq UltraScale+ MPSoC architecture, which features a quad-core Arm Cortex A53 processor, more than 250 thousand logic cells, and a H.264/265 video codec. The SOM also features 4GB of DDR4 memory and 245 IOs, which allow it to adapt to virtually any sensor or interface.” The company added that 1.4 tera-ops of AI compute allows developers leveraging the K26 to create vision AI applications offering more than three-times higher performance at lower latency and power compared to GPU-based SoMs.

Aiming for Vision AI

Targeting vision AI applications first makes a lot of sense, according to industry analysts, because they come with a high degree of development complexity and requirements for large amounts of processing from numerous sensors to be done at the edge.

Kutleng, a company with a smart camera platform for use in tracking cameras and other applications, already has used the Kria SoM to achieve faster, cheaper development of tracking camera solutions used for wildlife safety in national parks in South Africa. Benjamin Hector Hlophe, Director of Technology Operations, Kutleng, said in an email, “The Kria SoM is very well priced and allows for products that could not be done traditionally using other SoM board vendors. Xilinx has excellent documentation and reference design. This was a big added bonus and means we don’t need to do a custom board support package.”

Hlophe added, “We are under price pressure from others developing camera systems using competing solutions such as Nvidia Jetson/Xavier or Qualcomm GPUs/SoCs. Xilinx providing the Kria SoM with a rich set of vision libraries enables us to use our existing skills to develop highly differentiated products with superior performance, while also being price competitive.”

Jim McGregor, principal analyst at Tirias Research, told Fierce Electronics via email that the Kria K26 coil give many types of vision AI developers a leg up.“A big shift in AI is doing more of the neural network inference processing at the edge. This is driving the evolution of the hardware and software for processing at the edge, but not necessarily the need for a SoM. The benefit of the SoM is that you have a platform that works right out of the box with tools and pre-trained models. You even have an ecosystem that can customize the SoM to your form factor or performance requirements.”

Dave Altavilla, principal analyst at HotTech Vision and Analysis, noted in an email to Fierce Electronics, “Xilinx SOMs will allow developers to code in native frameworks and design environments and then deploy to a production-ready, ruggedized processor module with a robust Arm core CPU complex, configurable logic, 4GB of DDR 4 RAM, and a plethora of common IO interfaces from MIPI and LVDS to multi-gig Ethernet. Xilinx is also making IO carrier cards available as reference designs. So, all of the tools are there for customers, including Xilinx’s Vitis software that lets engineers target their design with their own differentiated technologies, all from the comfort of their preferred development environment with key library support.”

In addition to all of that, Xilinx is trying to demolish any last hurdles developers might face by also offering the KV260 Vision AI Starter Kit to help clients plan out their applications even before they start working with the K26 SoM. Finally, Xilinx also makeis making SoM-based apps available in a new embedded app store for edge applications.

“A critical part of deploying Ai is making it easier to use in terms of tools, libraries, and even pre-trained neural network models,” McGregor said. “The hardware really means nothing without the software and tools. So, creating an apps store makes perfect sense to create a platform that is easy to use and customize. Just think of how useful a smartphone would be without the millions of mobile apps.”

Entering a New Market

The latest move by Xilinx comes at an interesting time in the company’s history. The company is in the midst of being acquired by AMD in a $35 billion stock deal set to close later this year. Xilinx also announced Alveo accelerator cards targeted at AI and other types of data center workloads, which was seen as a step beyond its traditional chip business. The SoM announcement is the latest step in a new direction.

“Xilinx’s entrance into the burgeoning SoM market builds on our evolution beyond the chip-level business that began with our Alveo boards for the data center and continues with the introduction of complete board-level solutions for embedded systems,” said Kirk Saban, vice president, Product and Platform Marketing at Xilinx, in a statement. “The Kria SoM portfolio expands our market reach into more edge applications and will make the power of adaptable hardware accessible to millions of software and AI developers.”

Taking the plunge into the SoM market puts Xilinx in a rapidly expanding new market environment, but also a fragmented one. Xilinx described it as a market that is experiencing a compound annual growth rate of about 11%, on the way to being worth $2.3 billion by 2025.

“The System on Module market is a rapidly-expanding segment of embedded system developers, researchers and even the maker community,” said Altavilla. “At this point in time, Xilinx’s primary competition in the AI, machine vision and robotics markets are GPU-based solutions. However, the Kria K26 SoM differentiates by enabling not only programmability and dedicated compute resources, but powerful hardware reconfigurability and adaptability that only programmable logic solutions like FGPAs can offer. It’s perfect for AI inference processing at the edge in a compact, low power solution that can adapt as future needs evolve.”

RELATED: AMD to buy Xilinx for $35 billion in stock