Satellite constellations with extra onboard compute are several steps away from handling full AI workloads for people on Earth.
Key Takeaways:
AI data centers need vast amounts of electricity, water, and space, but that space doesn’t have to be terrestrial or even have an atmosphere.
Launching servers into orbit skirts some thorny obstacles for data centers, including:
As a result, increasing amounts of compute are being attached to satellite constellations or purpose-built orbital units.
“We call it data centers in space, but it’s going to be compute centers in space,” said Helmut Puchner, fellow and vice president of aerospace and defense at Infineon Technologies. “Because AI installation and AI data centers outgrow the speed at which we can supply enough energy and resources on the ground, we want to go to space. We’ve seen the announcements from SpaceX about its Terafab to build their own chips, and Blue Origin’s TeraWave satellite network. Tera seems to be the new word. SpaceX is serious about data centers in space. It’s very clear what goal they have. They announced a million satellites because they want to get 100 gigawatts of installed compute power in space. If you look at a satellite right now, experts tell me they max out at about 40 kilowatts per satellite. Maybe they want to push it to a higher level, such as 100 kilowatts. When multiplied by a million, that’s 100 gigawatts.”
Still, there are numerous technical challenges that need to be addressed. “The biggest obstacles are ample power generation, thermal management, long-term reliability in a radiation-heavy environment, and growing levels of space debris,” said Steven Woo, fellow and distinguished inventor at Rambus. “On Earth, we lean on airflow, liquid cooling, and easy serviceability, none of which translate well to data centers in orbit. Radiation effects on advanced nodes and memories add another layer of complexity that demands hardened designs and redundancy.”
Others agreed that the term ‘data center in space’ is a misnomer. “It’s more like a souped-up satellite, or a constellation of satellites,” said Pradeep Shenoy, compute power technologist at Texas Instruments. “If you look at what companies have released, they’re going to increase the power levels and the compute capabilities of the satellite, but they’re going to deploy a million of them to get to the compute levels they need.”
Whether this is a goal the industry wants to achieve is uncertain. “Knowing what it takes to power the current generation of chips to really, truly run AI, it scares me a little how big they’re going to have to build a solar collector to power all of that,” said Shawn Carpenter, program director for 5G/6G and space at Synopsys. “It feels like a pretty heavy lift. I can see how attractive it is to put things up there because of how cold it is, and how well you could keep things cool when you’re in the shadow side of things. But at the same time, you’re depending on the ability to collect enough solar to create the energy you need. How large does that collector need to be? At what point do astronomers on the ground really begin to complain because of the size of some of these arrays? We’ll have to see.”
There are also communication challenges because fiber-optic cables aren’t feasible in orbit, so large amounts of data needs to be transmitted wirelessly. “This requires spectrum management,” said Lang Lin, director of product management at Synopsys. “There could be interruptions due to electromagnetic pulses or other things that can disrupt communication. Maintenance is hard, too. On Earth, you can unplug a cable, replace a rack, or replace a computer. But in outer space, if something blows up, it’s not as cheap as going to your local data center and replacing something. It’s possible, of course, but it’s expensive.”
Removing heat in space
While electronics perform better in cooler temperatures, heat dissipation is a central engineering challenge for computing in space. “For 100 Nvidia GPUs, you would need 33 square meters of solar and 16 square meters of radiation panels to radiate the heat out,” said Puchner. [1] “You can do the thermodynamics calculations all day long, but it doesn’t matter. It’s the same situation. Those compute centers are basically just burning energy and converting it into heat.”
Forced liquid cooling, as implemented by data centers on Earth, will essentially work the same in space, but the heat must be transferred to radiators to vent out into space. “The convective cooling of air is replaced with liquid convective cooling, which is more efficient,” said Jeremy McCaslin, senior director of fluids product management at Synopsys. “You can still rely on conduction and radiation in space. Heat is conducted away from chips or from the hot to cold side through solid materials into heat pipes or fluid loops. The heat is then ultimately dumped into space via radiation, using, for example, large radiator panels. These radiators emit the heat as infrared radiation into space.”
Reducing heat by optimizing chip architecture
Like on Earth, energy-efficient chips in space can reduce heat generation at the source. “What’s unique and weirdly unexpected is how difficult it is to remove heat in space,” said Matthew Bubis, director of product management at Imagination Technologies. “But if your chips are more efficient, you’re generating less heat, and it’s easier to remove. That means you need fewer radiators for heat removal and heat exchange systems up there.”
Accelerated chip architectures purpose-built for geospatial intelligence and autonomous space operations will need to leverage advances in edge AI to eke out more performance in size, weight, and power (SWaP)-constrained environments. Radiation-hardened, automotive-grade edge AI GPUs can be a good fit for orbital compute and memory.
“Even before customers get into radiation hardening — and there’s only so much rad hardening that you can do — the IP can provide traditional safety architecture features,” said Bubis. “Techniques like error correction on memories or test loops inside the hardware make sure they’re continuing to correctly compute their inputs and correctly provide their outputs. We see a lot of these safety mechanisms, which are traditionally used for automotive advanced driver assist systems, being evaluated for space.”
NASA recently announced rad-hard, high-reliability, High-Performance Spaceflight Computing processors co-developed with Microchip, which use advanced Ethernet to connect multiple sensors or cluster several chips to enable spacecraft to process massive amounts of data onboard and autonomously make real-time decisions. [2]
Further, Nvidia recently announced dedicated platforms to bring data-center performance to space for companies, including Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space, and Starcloud. [3]
Ultimately, technology and monetary costs need to be considered. “The foundation of the idea of data in centers in space is economics and the unbelievable costs in both energy and infrastructure on Earth,” said Bubis. “But the free and continuous power you get in space with solar needs to be offset with the cost of launching and actually getting it up there. It comes down to the weight of the things that you’re taking up, and the power consumption of the compute systems that you have up there, and then also what you do about things like radiation.”
Sun-synchronous power supply
If compute centers are solar-powered, they need to be constantly synchronized with the sun at the North or South Pole, compared to regular satellites that move up and down between the poles, and often are powered down in between.
Constellations do not typically cover the poles. “If you look at the main constellations, they’re usually flying up to the 15th or 20th latitude,” said Infineon’s Puchner. “They’re not addressing the coverage in the polar regions, though there is a Terrestar constellation in Canada that’s going to be flying in that region, as well. For compute centers in space, you have a problem because you either have to have massive batteries, or you have to fly in sun-synchronous orbit, which is basically going around the North Pole, South Pole, North Pole, South Pole, so you have continuous access to the sun because you need the power and the energy continuously. You cannot buffer the energy. In a regular constellation satellite, when they fly over the ocean, where there’s not much activity, they power down the satellite. The satellite takes 90 minutes to circumvent Earth, then when it hits a populated area, they power it up again.”
Temporarily powering down won’t work for ODCs. “The data centers will still rely on the regular constellations for the upload and download of the data,” said Puchner. “Let’s say you enter a ChatGPT request and it’s calculated in real-time in a compute center in space. The latency it takes for you to get the data back is the key. It requires the sun-synchronous compute center, plus low-latency access globally. I’m not sure how they’re going to manage that, because right now the data transfer for the last hop is in the mesh with a standard constellation.”
Fig. 1: Global satellite communication. Source: Infineon
The multi-hop solution is likely to dominate until other solutions can be found. “You are likely to have a system of distributing data flows around the earth, compared to where the solar power is and where the compute is actually being run, which would need to be in a stationary orbit that has continuous power,” noted Imagination’s Bubis.
For example, Axiom Space’s satellites collect raw data, such as images or telemetry data, then send it to a nearby orbital data center node via 2.5 Gbps-capable Optical Intersatellite Links (OISLs), with 10 Gbps on the horizon. The ODC runs processing and inferencing, including filtering images, detecting features, compressing files, or running AI/ML models, before data is sent back to Earth via commercial relay constellations in Low and Geostationary Earth Orbit. [4] In addition to ODCs, Axiom Space installed a data center node on the International Space Station [5] and is actively building a commercial space station. [6]
Synopsys’ Lin noted the possibility of powering ODCs with nuclear energy, along with solar and batteries. While Radioisotope Thermoelectric Generators (RTGs) have been used to power deep space probes, nuclear-powered ODCs face hurdles. [7]
Simulating sun-synchronous positioning
Synopsys recently partnered with Cesium to support NASA’s rollout of a lunar network. [8] The company takes customers’ design data, MCAD models, reality capture, and trained data and converts it into open standard 3D Tiles, which enables the performance streaming of disparate massive 3D geospatial data sets at runtime, explained Alex Paulson director of business development at Cesium. The resultant model is 4D, because it can show a time lapse of the 3D model through the day as it moves.
“We can time-sync it in our mission simulation software,” said Paulson. “You can go to any specific moment in universal time and see the exact orientation of all of the planetary orbs, satellites, moons, and all that stuff going on.”
For satellites, spacecraft, and orbital data centers that need solar power, the model helps understand when their solar panels are going to be picking up sunlight. “Then I can know how to time my operations according to when I have peak power,” said Synopsys’ Carpenter. “When I don’t have solar illumination, I don’t get power. I must be careful to power down or protect certain systems on a satellite. We can do that in our modeling and simulation of the kinetic motion of things through the availability of 4D models.”
Memory and reliability in advanced chips
When the solar arrays are synchronous with the sun, components can be shielded from the heat — but memory is still vulnerable.
“The electronics are protected and sitting in an encapsulated, shielded environment,” said Infineon’s Puchner. “All these thermodynamic challenges need to be solved, but it’s not a big issue. One problem is, if you fly in sun-synchronous orbit, you’re going to the polar regions, and you have a high flux of protons from the magnetic field being funneled into the pulse; the radiation environment is much higher. That’s another thing to watch for in the industry. A lot of companies have claimed to fly Nvidia equipment in space, but none have reported success so far. It is easy to fly, but did it work?”
While radiation tests have been done on Nvidia equipment, memory behavior at the poles remains an ongoing variable. “The critical portion of their equipment is the memories, such as classical DRAM, DDR5, or HBM memories, as well as the boot code memory, which is a commercial NorFlash,” said Puchner. “All of these are vulnerable to radiation offsets. If you get hit and the system hangs up, or if the hit rate is so fast that you cannot get any decent calculations out, it’s useless. We have to see what the survival rate is. In a regular 500- to 600-kilometer orbit close to the equator, it’s not a big deal. You might survive, and they are surviving — SpaceX Starlink constellations, Amazon Leo, and so forth. But for sun-synchronous memory, it might be different.”
Fig. 2: A solar-powered satellite. Source: Infineon
Advanced electronics are more expensive, and triple redundancy means triple the cost.
While radiation can be managed, single-event upsets are more variable. “How do you deal with the real threat of single events with a COTS (Commercial Off-The-Shelf) technology approach, that is, you’re buying your advanced processors in unhardened TSMC or Intel, really advanced processes, and now they need to be hardened, or at least tolerant?” said Hugh Barnaby, leader of extreme environment reliability [9] in the U.S. Department of Defense’s ME Commons SWAP Hub at Arizona State University. “In the old days, NASA and JPL (Jet Propulsion Lab) didn’t care because they weren’t building hugely complex systems. They were very efficient, and they didn’t need to buy that many parts.”
Spending $20,000 to $100,000 — or more — on a sophisticated component, such as a finFET (fin field-effect transistor) or gate all around (GAA) IC, is manageable if systems only require a limited amount. “But if you go out and start building very large data centers or compute centers in space, you’re going to be buying and moving a lot of parts,” said Barnaby. “The only way to deal, for sure, with single event upsets in the memories of in finFET and GAA is to put in triple-mode redundancy. That’s very expensive.”
Reducing the costs of making a robust suite of electronic components that work in the Earth’s magnetosphere is the real challenge. “SpaceX and high-tech companies are all about trying to stay cheap on that and taking the risks,” said Barnaby. “They’re not nearly as cautious as NASA. I don’t think they’ve had any incredibly bad failures [up in space] yet, so at this point they think that the risk has been worth it.”
To date, most of SpaceX’s failures have involved launching its rockets, which are developed with a “fail fast, learn fast” philosophy.
Conclusion: the near future outlook
Orbital data centers already exist in the form of additional compute on satellites to filter and process space data locally. Larger-scale data centers that resemble terrestrial units, and which are used primarily for training or inference of AI models used by humans on Earth, are not quite here yet.
“While space-based computing is indeed possible, the economics and reliability challenges will prevent it from being mainstream in the near term,” said Rambus’ Woo.
For example, Google is reportedly in talks to use SpaceX to launch space data centers [10], but Amazon’s Jeff Bezos said that Elon Musk’s timeline of two to three years is a bit too ambitious. [11]
“Big organizations want to have compute centers and data centers in space because they don’t want to deal with Earth’s limitations, regulations, permits, and energy,” said Infineon’s Puchner. “Local power companies are not ready with the power upgrades to run a data center, but the data center is ready. It’s too much frustration. The projected growth of AI in our environment, in our work life, in our daily life, is going up, so demand is there, and business is there, but it’s too slow. So in the next five years, we will have a significant amount of compute power installed in space.”
There’s a lot of interest and investment in it, agreed TI’s Shenoy. “The argument is that it solves some of the energy supply challenges in theory, because you’ve got the sun, and if you are in sun-synchronous orbit, you can leverage energy from the sun and not be constrained by power needs. In that regard, space is a very harsh environment, and depending on the level of orbit you’re in determines what sort of radiation is going to happen. It’s non-trivial, and this does not consider the cost it takes to send something into space. There are a lot of questions about how to maintain compute capabilities in space. With hot swap circuits in Earth-based data centers, if something happens with one of the devices in this tray, you need to be able to do the maintenance. There’s an open question of how to do that in an effective and capital-efficient manner.”
Fabs in space could be next. “On the manufacturing side, people say they’re literally going to go up and fabricate with cleanrooms in space,” said Barnaby. “I give that a ‘soft maybe.’ The advantage is that the low-pressure vacuum environment helps. Clean rooms have a lot of evacuated spaces where you want to have things that way. It helps with depositions and things like that. It should be much cleaner, so the contamination would be lower. Those two things are really advantageous. The downside is you’ve got to move stuff back and forth, and you have to fix things when they break. That is where radiation and temperature are part of the mix. It would need to be big enough to house the radiators for the deposition. The power needs would be pretty significant, too.”
[Editor’s note: Future articles will cover lunar networks, security challenges in space, and ICs in extreme environments.]
References
Related Articles
Data Centers In Space?
Someday, but not soon. Chip execs don’t need to start designing for space just yet.
Extra Safety Measures Needed For Aerospace ICs
Fail-safe operations and boot load memory are essential for ensuring pricey resources aren’t lost in space.
Radiation, Temperature, Power Challenges For Chips In Space
Semiconductors designed for spacecraft face a range of extreme conditions that impact aging. A reliable source of power is also needed.
Security Risks Mount For Aerospace, Defense Applications
Supply chain vulnerabilities, hardware attacks, and communications hacks are rife. Autonomous technology poses extra threats.
 |
 |
 |
 |
 |
 |
 |
 |
Leave a Reply