Xilinx launches Kria chips to handle AI for edge applications

Xilinx has introduced its Kria programmable chips and boards for holding AI applications at the edge of the network. This should come in handy for visual applications like smarter cameras.

San Jose, California-based Xilinx, which is in the process of being acquired by Advanced Micro Devices (AMD) for $35 billion, has a group of products dubbed the Kria portfolio of adaptive system-on-module offerings for AI at the edge.

These are production-ready small form factor embedded boards that enable rapid deployment in edge-based applications. Coupled with a complete software stack and prebuilt, production-grade accelerated applications, Kria adaptive modules are a new method of bringing adaptive computing to AI and software developers.

The first product available in the Kria SOM portfolio, the Kria K26 SOM, specifically targets vision AI applications in smart cities and smart factories. The Xilinx SOM roadmap includes multiple products, from cost-optimized SOMs for size- and cost-constrained applications to higher-performance modules that will offer developers more real-time compute capability per watt.

Xilinx said industry reports put market growth at 11% per year and total market revenue at $2.3 billion by 2025.

“I think there’s solid potential here,” HotTech Vision and Analysis analyst Dave Altavilla said in an email to VentureBeat. “Similar to the way Nvidia enabled its graphics processing unit (GPU)-powered edge AI solutions with Jetson and Jetson Nano, Xilinx’s Kria SOM is a complete shrink-wrapped solution for developers, engineers, and makers to get their feet wet with adaptable field-programmable gate array (FPGA)-based acceleration, with little to no experience required in programmable logic (FGPAs).”

He added, “In addition, Xilinx’s embedded app store for edge AI will help foster accelerated solutions with faster time to market in anything from facial recognition to natural language processing applications and more. In short, it’s what the company needs to demystify the FGPA for the masses, and in conjunction with Xilinx Vitis software development tools, will allow engineers to work in their own native frameworks and programming languages like PyTorch, TensorFlow, Caffee, C++, OpenCL, and Python.”

He added, “It’s a clear indication that the adaptive nature of FPGAs doesn’t need to be relegated to just the power user-programmable logic engineer anymore. And with Ubuntu support on the way, these dev kits could go mainstream in a hurry.”

Faster to deploy

Above: Kira Kria SOM module.

Image Credit: Xilinx

Kria SOMs use Xilinx adaptable hardware, delivered as production-deployable, adaptive modules. Kria SOMs can be deployed rapidly using end-to-end board-level solutions with a prebuilt software stack. By allowing developers to start at a more evolved point in the design cycle compared to chip-down design, Kria SOMs can reduce time to deployment by up to nine months, Xilinx said.

The Kria K26 SOM is built on top of the Zynq UltraScale+ MPSoC architecture, which features a quad-core ARM Cortex A53 processor, more than 250,000 logic cells, and an H.264/265 video codec. The SOM also features 4GB of DDR4 memory and 245 input-output paths that allow it to adapt to virtually any sensor or interface.

With 1.4 teraflops of AI compute, the Kria K26 SOM enables developers to create vision AI applications offering more than 3 times higher performance at lower latency and power compared to GPU-based SOMs. This is critical for smart vision applications, including security cameras, city cameras, traffic cameras, retail analytics, machine vision, and vision-guided robotics.

Xilinx has invested heavily in its tool flows to make adaptive computing more accessible to AI and software developers who lack hardware expertise. The Kria SOM portfolio takes this accessibility to the next level by coupling the hardware and software platform with production-ready vision accelerated applications.

Lots of applications

Above: Side view of the Kria module.

Image Credit: Xilinx

These applications eliminate all the FPGA hardware design work and only require software developers to integrate their custom AI models and application code. They can optionally modify the vision pipeline — using their familiar design environments, such as TensorFlow, Pytorch, or Café frameworks, as well as C, C++, OpenCL, and Python programming languages — enabled by the Vitis unified software development platform and libraries.

Xilinx offerings are open source accelerated applications, provided at no charge, and range from smart camera tracking and face detection to natural language processing with smart vision.

The Kria KV260 Vision AI starter kit is priced at a low $199. When customers are ready to move to deployment, they can seamlessly transition to the Kria K26 production SOM, including commercial and industrial variants priced at $250 or $350, respectively.

The KV260 Vision Starter Kit is available immediately, with the commercial-grade Kria K26 SOM shipping in May of 2021 and the industrial-grade K26 SOM shipping this summer. Ubuntu Linux on Kria K26 SOMs is expected to be available in July.