This week Intel dished on putting its first AI/vision processing chip into orbit onboard an European Space Agency (ESA) cubesat last month. This demonstration is the first of several planned and is part of a larger trend to put more compute power into orbit to process raw data from imaging satellites.
Launched on September 2, ESA’s PhiSat-1 contains a hyperspectral-thermal camera to monitor polar ice and soil moisture. The 6U CubeSat is roughly the size of a large cereal box and includes an Intel Movidius Myrad 2 Vision Processing Unite (VPU) to filter through the mass of data generated by the camera. The low-power Myrad 2 was pounded by 36 hours of high-speed radiation beams at the CERN labs in Switzerland to assess its ruggedness for space applications.
Researchers want to look at ice and soil, but about two-thirds of the planet’s surface is covered by clouds at any given time. Downloading all the raw data to Earth and then having a scientist or ground-based AI wastes precious downlink bandwidth and review time.
“The capability that sensors have to produce data increases by a factor of 100 every generation, while our capabilities to download data are increasing, but only by a factor of three, four, five per generation,” says Gianluca Furano, ESA data systems and onboard computing lead. “And artificial intelligence at the edge came to rescue us, the cavalry in the Western movie.” Using the onboard processing to identify and discard cloudy images should save about 30 percent of bandwidth, plus staff time not spent looking at irrelevant pictures.
PhiSat-1 is the first of several planned Intel VPU projects. Intel is working with its vision integration partner Ubotica to put Movidius Myriad X into a space vehicle launch camera and next year intends to deploy MyriadX hardware and demonstration applications on the International Space Station in collaboration with NASA’s Jet Propulsion Lab.
Ubotica believes more edge computing being embedded into satellites over flying data centers, at least for smaller cubesats designed to be quickly manufactured and launched at lower cost.
“We see three major factors currently driving the demand for onboard processing on satellite: latency reduction, autonomy, and data filtering (reducing the amount of data to downlink),” said Aubrey Dunne, CTO of Ubotica. “In a terrestrial application all of these factors may be most easily addressed by increasing compute capabilities. However space has some unique factors that can make the addition of extra compute more difficult: power consumption, mass (and volume), and radiation effects. The first two of these in particular are important for cubesats, where the current drive is to reduce build and launch costs. So in short, we do not see a move to ‘servers’ for cubesats, but rather a continued integration of commercial grade processors. As commercial grade silicon performance improves per unit power/mass/volume, cubesats processing capability will likely follow.”
Other companies have alluded to putting specialized processors onboard their satellites for in-orbit data processing. Capella Space has incorporated a NVIDIA processor into its synthetic aperture radar (SAR) satellites to conduct some quick-and-dirty image processing so users can get a rough look at an image while a larger data file is awaiting download and processing on the ground.