Xilinx has launched its first FPGA with space qualification for a decade, designed to support machine learning and artifical intelligence frameworks.
The XQRKU060 Ultrascale Radiation Tolerant (RT) Kintex FPGA is built on a 20nm process, compared to the 65nm process used for previous space qualified Virtex devices. This gives four times the logic capacity, with 633 logic blocks, and ten times the digital signal processing with 2760 DSP slices, as well as 32 Serialiser/deserialiser (SERDES) high speed interfaces.
“Machine learning in space is in its nascent stage but starting to make ground in the next four to five years,” said Minal Sawant, Space Systems Architect at Xilinx. THe move to 20nm enables up to 5.7 teraoperations per second (TOPs) of peak performance for 8bit integer frameworks (INT8), nearly 25 times that of the previous generation.
For machine learning, the architecture has been extended with wide 27 x 18 multiplier and extra accumulator blocks which otherwise would have to be implemented in the logic array. Initial support is for open source machine learning compilers and DNN array such as TensorFlow and PyTorch, but d the future roadmap includes the Xilinx Vitis AI framework. “This will open up a whole new world of how to use machine learning in space,” said Sawant.
One initial use for the device is in large geostationary satellites that handle communication links, hence the 32 high speed SERDES interfaces. Another is for pattern recognition of camera images.
A key capability is that the parts can be upgraded in space, for example to upgrade an ML inference framework, However the initial use is likely to be bug fixes, says Sawant.
The FPGA can be reprogrammed by changing the 192Mbit bitfile that configures the chip. With larger 1Gbit non-volatile memories qualified for space use, up to three separate bitfiles can be stored with completely new functions.
“The Ultrascale Kintex has true unlimited in-orbit reconfigurable