These types of chips, which can endure harsh radiation in space, could be the brains of orbital space applications such as satellite payloads, where they offer a 10-fold increase in digital signal processing performance over previous versions, Xilinx said. That means the chips can take data from sensors and process them efficiently.
Xilinx designs field programmable gate arrays (FPGAs), or chips that can be programmed or re-programmed after they’re put into hardware. The Kintex UltraScale FPGA chips are the first that use a 20-nanometer manufacturing process (a nanometer is a billionth of a meter), replacing the older 65-nanometer process previous chips used.
The XQRKU060 chip also brings high-performance machine learning (ML) to space for the first time. A diverse portfolio of ML development tools, supporting industry standard frameworks that include TensorFlow and PyTorch, enables neural network inference acceleration for real-time on-board processing in space with a complete “process and analyze” solution.
The chip has dense, power-efficient computing with scalable precision and a large on-chip memory. It provides 5.7 tera operations per second (TOPs) of peak INT8 performance optimized for deep learning. That’s 25 times better than the prior generation.
The result is more processing for a significant reduction in size, weight, and power — all key factors when you’re designing a chip that is used in the harsh environment of space where physical space is a premium. Minal Sawant, system architect and space products manager at Xilinx, said in a statement that the tech is “ideal for high bandwidth payloads, space exploration, and research missions.”
The XQRKU060 is the industry’s only on-orbit reconfigurable chip, Xilinx said. That means it could be up in space in a piece of hardware and then be reprogrammed to do something else.
The on-orbit reconfiguration capabilities, together with real-time on-board processing and ML acceleration, allows satellites to update in real time, deliver video-on-demand, and perform compute “on-the-fly” to process complex algorithms. The company said the ML capabilities suit a variety of problems spanning scientific analysis, object detection, and image classification, such as cloud detection.
Flight units of the 20-nanometer RT Kintex UltraScale space-grade XQRKU060-1CNA1509 FPGA will be available in September.