Specialized in the vision and analysis of images, these specialized, energy-efficient chips are one of the technological building blocks that will activate the automation of space exploration. And make possible a future ever closer to science fiction.
While the power of the chips in our computers and smartphones has seen phenomenal progress over the past decade, there is one world that is watching this deluge of power with envy: the space industry. In a context of a space conquest rich in projects – constellations of satellites, lunar and Martian projects, CubeSats, etc. – the electronic reality of the machines that we send into space is very different from what we put in our phones. Space is hostile territory for semiconductors, and the performance of chips reinforced to withstand the radiation of the great vacuum is far from those that operate on dry land.
But things are moving and a chip could accelerate our conquest and exploitation of space: the VPU. This image processor (Visual Processing Unit), we have seen it develop in particular in drones, autonomous cars and other video surveillance systems. Their role ? Interpret information from cameras faster and more efficiently than CPUs and GPUs. In this very new world of VPUs, a tiny, one-square-centimeter chip is making its way into the stars: Intel’s Myriad processor.
The data and calculation puzzle
In space, no one can hear you scream. But there are certainly not many people to take charge of your calculations. While they can properly manage their machines, probes and other cameras, space-qualified processors are not models of power. “ So far, all space missions (involving imagery, ndr) happen in the same way: the satellite sends the data to earth and it is the ground equipment that does the calculations »Explains Gianluca Furano, engineer specializing in embedded computer systems for ESA, the European space agency.
The problem ? The bandwidth to send data from space to the ground is limited. Boosting image processing or throughput leads to an increase in energy consumption and heat dissipation, two extremely limiting factors in space. In a classic scenario, the satellite slowly sends all of its images, some of which are of no interest to the application (clouds, ocean areas, etc.).
This is where Myriad 2 comes into play. Far from being the latest in silicon, this chip initially intended for drones works wonders in space. ” It’s a real revolution that will take place in space exploitation “Enthuses Gianluca Furano, whose first CubeSat, PhiSat-1 (or Φ-Sat-1 for nerds), was successfully launched on September 2. A satellite whose mission is purely technical: dedicated to imagery with its hyperspectral camera, it must demonstrate the relevance of the use of AI chips in space.
ESA – Capable of understanding the content of images, the Myriad 2 can not only sort what needs to be sent, but also send not the raw images but the processed information – here, the geographic coordinates where triggered forest fire.
If it were conventionally equipped, this nanosatellite would have no other choice than to transfer the images drop by drop to a terrestrial relay. Slowing down the processing of information sometimes by several hours. ” With the Myriad 2 chip, rather than sending the raw data, the satellite takes care of it itself and only sends us the analyzed data. For example, it can detect a forest fire and send us the information and GPS coordinates directly. While it sometimes took up to a day to transmit, process and compile information, with a VPU the process now only takes a few minutes! », He rejoices.
Lots of power in just 1 watt
The strong point of VPUs is their specialization which makes them very energy efficient. A blunt argument in space. “ We did not blindly integrate the Myriad 2, we carried out numerous benchmarks, numerous internal measurements. From fPGAs to CPUs, spatial or not, to PC GPUs or even to Nvidia’s Jetson, etc. What we found is that in image analysis Myriad 2 performs just as well as Jetson, but with only 1 watt of TDP! Jetson alone dissipates 10W, that is to say the total envelope of the electronics of PhSat-1! “. However, if on earth we can always add fans, heat pipes, watercooling, to obtain the best performance, space does not allow this kind of follies.
” Up to 5W, you don’t need active cooling solutions. But after this threshold, you have to start thinking about it, which adds weight, bulk, etc. »Mr Furano explains. ” To give you an idea, adding an instrument that consumes 1 watt on a Martian rover means adding a little more solar panel here, a little more power there, which inflates mass to take away. At the end of the chain, 1 watt to supply and dissipate in addition is 1 ton of fuel to burn more! “
The energy efficiency of a VPU like the Myriad, which is a king of inference, therefore allows savings in terms of energy, bandwidth and mass to be sent since it does not need to ‘to be cooled. But it also lowers the hardware and software bill. ” As Intel’s chip has been produced in millions of copies, it is inexpensive compared to space components and it takes advantage of good software tools “. What to lower the bill and the working time. Which doesn’t mean that everything was easy.
Space, a hostile environment for fleas
SSA / ESA – Electromagnetic radiation from space is a constant threat to electronic devices as well as to living things.
” It took us two years to qualify our material base for space », Explains G. Furano. ” We have reinforced the electronics surrounding the processor, especially the power supply “. But as Myriad 2 is not a component developed for space, we had to develop a software part that takes into account the constraints. Thus, the very operation of memory management and access has been changed. Impossible to leave elements in the RAM, the space radiation can falsify the data which is stored there even put the system out of service.
CERN / Maximilien BRICE – A team from ESA and the Ubotica company used CERN’s particle accelerator, the Super Proton Synchrotron (SPS), to test its resistance to radiation.
The solution ? Always start from a blank sheet: ” To avoid errors caused by the accumulation of tasks, the memory is completely emptied between each image calculation. We have designed algorithms and software to control the data stored in memory to guarantee their reliability. And redundancy routines in case parts of the chip do not work properly “.
Adrian BRANCO / 01net.com – The current chip called Myriad X (here in its consumer version integrated into the “Intel Neural Compute Stick 2” USB key) is up to x10 more efficient than the Myriad 2 with which the ESA works .
In total, it took the European team four years of work to develop the CubeSat PhiSat-1. This explains why it does not integrate the Myriad X launched in 2018, but the “old” Myriad 2 from 2016. “ Spatial development takes time and we will never be able to keep up with the speed of consumer product development “, Explains Mr. Furano. But even soon to be two generations behind – Keem Bay, the replacement for Myriad X should soon be on the market – this VPU is a revolution.
VPUs (a little) open the door to science fiction
Tim Herman / Intel Corporation – Scientists at the Tyvak International Operations Center in Turin, Italy, monitor the satellites.
” VPUs open up new horizons in observational science, for both fires and agriculture. But they will also make possible what is today infeasible Mr. Furano exclaims. Among these impossible tasks? The docking of a machine. ” A meeting with a non-cooperative object (like a drifting satellite, ndr) is absolutely impossible without a stand-alone visual computing system.
Tasks where the Myriad seems to be a champion. ” Training machines is important, but it represents 1% of tasks. The most important for embedded systems is inference, which represents 99% of the needs “. In short: the power required to drive the machines is only important during the algorithm creation phase. Once these have been developed, the needs are mainly turned towards fast and energy efficient chips – which explains, again, the choice of a Myriad chip compared to an Nvidia Jetson, more efficient in learning, but less in terms of application.
From automated piloting systems to local data processing, VPUs thrill scientists, researchers and space engineers. Who use components with great pleasure… which are not intended for them. ” Space can never lead to large production volumes », Mr Furano readily admits. But the community relies on another industry: the automobile. With the advent of the autonomous car, the sensors and processors involved in autonomous navigation systems are gaining momentum. ” It is clear that (in the future) we will be using a lot of the components that are (integrated) in self-driving cars. »He continues.
A success that leads to others
ESA / CNES / Arianespace / CSG Video Optics – JM GUILLON – This is flight VV16 of the Vega rocket launched on September 2, 2020 which put 53 satellites into orbit including PhiSat-1.
Φ-Sat-1 is not a classic scientific mission, but a technical test which aimed to validate the relevance of VPUs in space uses… which is done. The Gianluca Furano teams have in the process validated two hardware platforms integrating Myriad-2, each with its own software framework suitable for small or large satellites.
And these design efforts, both hardware and software, will thus serve not one, but five missions, all of which will result from a public-private partnership. A success which allows Intel to sign a great success in low consumption IA chips, a success very different from that of the “thermal power stations” that are the server chips – EPYC / Instinc at AMD, A100 at Nvidia, Xeon and Xe HPC at Intel.
A paradigm shift for Intel, whose model was initially shaped by the race for the power of its x86 architecture. And which will perhaps allow the giant, heckled in its various sectors, to resume the technological leadership to which it was accustomed.