Xilinx is in the middle of an overt change in go-to-market strategies that is reaching into new markets. For many years, the Xilinx model was that it made FPGAs, and partners made the board, logic, and other bits to turn FPGAs into useful devices. Now, Xilinx is attempting to grow the TAM for its FPGAs by pushing its products into new areas. Xilinx has its Alveo accelerator boards for server PCIe slots and has expanded into building ready-to-deploy solutions. Now, the Xilinx Kria line is designed to be system-on-modules that will accelerate edge AI inferencing using hardware and software from Xilinx.
Xilinx Kria Edge AI SOMs Launched
First off, Xilinx Kria is a new brand name like Xilinx Alveo is. The goal of Kria is to provide system-on-modules (or SOMs) that can be quickly integrated into designs to accelerate AI inferencing.
The first module is the Kria K26. This is designed primarily as a vision accelerator. There are going to be other SOMs, but this the K26 is the first one that the company is launching today.
These have quad-core Arm Cortex-A53 cores along with Xilinx FPGA logic. One can see the interfaces below, but the key here is that some of this logic like the USB and video codes are hardened. That helps lower cost and also saves effort for designing around these modules.
Using a common form factor, the idea is that the basic SOM can be integrated into other devices quickly, and that lowers the time-to-market. One does not have to license or build a memory controller as another example.
With today’s announcement, we will have the Xilinx Kria KV260 starter kit which includes the Kria K26 SOM with a cooler and a baseboard. This is a $199 developer kit. The module is available in commercial and industrial grades at $250-$350. While the modules cost more than in the developer kit, the idea is that they are available in larger volumes and the developer kit’s goal is to onboard new users to Xilinx’s platform with rapid prototyping capabilities.
The Xilinx Kria KV260 starter kit can support up to 3 MIPI sensors and has display outputs (HDMI and DisplayPort) there is also 1GbE and USB onboard along with a way to get the SOM power. The module itself also has a cooler for desktop development.
Although this is not called out on the slide, this seems to be in response to NVIDIA Jetson which has adopted this SOM architecture. We just covered the NVIDIA Jetson TX2 NX that launched last week as that NX form factor can scale through a number of different SKU levels. The idea there is that a designer can build hardware and then choose the SOM that is right for the end product market’s needs without having to do entire redesigns of their platforms. Xilinx is doing the same thing.
Xilinx has gone down a similar route before. The Xilinx Pynq-Z1 was an attempt at a small development board for its Zynq FPGA.
The difference is that the Kria KV260 is designed to be a development platform around a deployable SOM. Xilinx has been working to get the unit up and running within an hour.
Xilinx is not just providing the FPGA. Instead, it is providing the FPGA with software stacks and features that are pre-built. It also will support Ubuntu in this generation, not just Yocto Linux. There will be Vitis AI solutions built to help this SOM be used with popular frameworks. Finally, one can program the programmable logic using RTL, but also higher-level tools. Xilinx is on a journey to lower the barrier to entry in using its products so software developers can start deploying on FPGAs without having to learn the intricate details.
Along this path, Xilinx has basic engines that will come with the unit. The idea is that companies and developers can use this starting point and customize as needed using various Vitis levels of abstraction.
At an even higher level, Xilinx has a number of applications that it will have in its library to help accelerate functions even faster. It is opening this up to partners to try getting an ecosystem behind the Kria SOM line.
As one would expect, Xilinx has a few comparisons showing that it can do more inferencing than less power than NVIDIA.
Of course, we will note Xilinx is comparing to GPUs, but it would have been nice if Xilinx just did a Kria to Jetson direct comparison.
Overall, what Xilinx is trying to accomplish is a high-risk/ high-reward endeavor. It is hard to jumpstart an ecosystem when NVIDIA has a dominant position. At the same time, the edge inferencing space is where NVIDIA faces perhaps the biggest competition right now and Xilinx has the resources of a large company behind it.
In other news, allegedly STH will be getting a Kria K26 developer kit, so we are going to see how that goes hopefully in the next week or so. Stay tuned to STH for more.