Artificial intelligence inference is about to become a lot more viable for makers of commercial space satellites thanks to Advanced Micro Devices Inc., which said today the latest version of its reprogrammable Versal system-on-a-chip is ready for takeoff.

More specifically, the company said that its XQR Versal AI Core chip has officially been granted a Class B qualification. Class B is a U.S. Department of Defense specification that determines if computer chips are capable of operating in outer space, providing an assessment of their performance, build quality, reliability and other capabilities.

Having achieved Class B qualification, the radiation-resistant Versal SoC is all set to start shipping by early next year. Among the interested customers is the U.S. defense contractor Raytheon Co., which said it will use the XQRVC190 as the basis of its next-generation space processors to power new satellites and spacecraft.

AMD acquired the Versal adaptive SoC business as part of its $49 billion deal to buy Xilinx Inc., a maker of field-programmable gate arrays. At the time, the company said it was buying Xilinx to compete better with Intel Corp. and Nvidia Corp. in the AI compute business.

The Versal SoCs should make it possible for satellites to process data and infer on AI with much greater performance than was previously possible, AMD said. That’s because it comes with capabilities that were previously available only with customized application-specific integrated circuits, which the company said are “prohibitively expensive” for many satellite operators.

It certainly packs a punch. The seven-nanometer SoC is powered by dual-core Arm Cortex A72 and dual-core Arm Cortex R5 embedded central processing units, more than 400 AI compute engines, 900,000 logic cells and 191 million bits of memory, AMD said.

What really makes the Versal SoCs so compelling, however, is that they can be reprogrammed on the fly to take on different kinds of workloads, AMD said. As a result, satellite operators will be able to switch their processing algorithms when required to ensure their communications and remote sensing applications are in tip-top condition. That’s not possible with ASICs, which incorporate fixed logic and cannot be altered once they’re built.

“Its heterogeneous computing capabilities and reconfigurable logic fabric will enable our teams to integrate more on-board processing in a considerably smaller footprint, enabling unprecedented advances in system-level size, weight and power,” said Barry Liu, Raytheon’s senior director of space systems.

Holger Mueller of Constellation Research Inc. said AMD is potentially opening up a new battleground to see who can power AI workloads in space. “It’s not possible for many chips to operate in space because of the exposure to radiation, which necessitates highly specialized silicon,” he said. “However, satellite-based data processing is extremely desirable, as there is only limited bandwidth available to connect these systems back to Earth. Similar to how software makers have attempted to bring their code closer to the data, chipmakers are now trying to put more capable CPUs in space, and AMD is taking a real step forward with its Versal AI Core SoCs.”

Image: PIRO4D/Pixabay