AMD takes AI inferencing to space with Versal chip

You can get further out on the edge than space. And that’s where the AMD Versal AI Edge chip is headed, said Advanced Micro Devices.

The new adaptive computing chip — the Versal™ AI Edge XQRVE2302 — is a radiation-tolerant, space-grade technology built as a complete system-on-chip (SoC). It’s qualified for space flight.

The chip stands out as the first adaptive SoC designed specifically for space applications in a compact form-factor, measuring just 23 millimeters squared. Compared to the existing Versal AI Core XQRVC1902, the XQRVE2302 boasts a significantly smaller board area and power savings of nearly 75%.

The XQRVE2302 integrates the enhanced AMD AI Engine (AIE) technology known as AIE-ML. This technology has been optimized for machine learning (ML) applications, offering extended support for prevalent data types in ML inferencing, such as INT4 and BFLOAT16.

The AIE-ML delivers superior performance compared to the original AIE, making the XQRVE2302 ideal for anomaly and image detection applications. With this capability, developers can efficiently convert raw sensor data into valuable insights.

The chip is a field programmable gate array (FPGAs), which can be easily reprogrammed in the field. The the XQR Versal adaptive SoCs offer unlimited reprogramming during development and even after deployment, including in-flight operations within the harsh radiation environment of space.

The Versal Adaptive SoC incorporates robust security features that prevent tampering and unwanted configuration changes. This empowers satellite operators to modify processing algorithms safely post-launch, facilitating flexibility in remote sensing and communications applications.

AMD said it has thoroughly tested the radiation tolerance of the XQR Versal SoC devices in collaboration with independent organizations. The results demonstrate that these devices are capable of supporting missions from low-earth orbit to geosynchronous earth orbit and beyond, ensuring their suitability for a wide range of space applications.

While commercial pre-production devices are currently available for interested customers, flight-qualified parts are anticipated to be ready for deployment by late 2024.