The dawn of a new space age and the promise of falling launch costs have many looking to shift compute off planet and into orbit.

At the extreme end, startups and tech billionaires alike are considering deploying data centers in space, a challenge that faces many hurdles. Irrespective of those lofty efforts, the traditional satellite industry is also embracing newer, more powerful semiconductors, with record amounts of compute filling the skies.

“It’s exponential,” AMD space systems architect Ken O'Neill tells DCD. “There’s so much compute going up.”

How low can you go?

O'Neill has spent the last 20 years developing hardware for the space sector, and says recent years have seen a marked difference. Thanks to SpaceX leading a dramatic decrease in launch costs, the number of satellites being put into orbit has skyrocketed, as have their individual requirements.

The majority of those launches now are for low Earth orbit (LEO) constellations serving global Internet efforts, most notably SpaceX’s own Starlink service. The proximity to the planet allows for lower latency and the ability to pick up weaker signals (including mobiles), but comes with an obvious disadvantage.

With the satellites so close, they lose line of sight to the end user quickly, necessitating the vast web of systems that Starlink and others are deploying. “So now you have to deal with things like cross-linking and the handover of the signal,” O’Neill says.

“In cellular communications on the ground, this was solved a few decades ago. But doing it in space, you have to have the computing horsepower to do those handoffs as well as the throughput capability.”

This additional compute needs to be in "a radiation tolerant form, and in such a way that they don't consume too much power and generate too much heat to cause massive thermal problems on the satellites."

In LEO, satellites face a barrage of radiation. "It's an environment that's very rich in protons," O'Neill says. "And protons can cause upsets in configuration registers, they can even cause latch-ups in certain integrated circuits."

The need to be more radiation tolerant has also pushed the industry towards newer hardware as, the smaller the process node, the lower the operating voltage.

"Reducing operating voltage makes you less susceptible to destructive effects," O'Neill explains. One issue, a single event latch up, sees the satellite conduct a lot of current from power to ground through the integrated circuit, potentially frying it.

"As you bring the operating voltage down, you pass a threshold beyond which you're not going to get single-event latch-ups." Modern integrated circuits are a lot less susceptible to these single-event latch-ups, but are not completely immune.

"While the core of the circuit may be operating at a very low voltage, 0.7 or 0.8 volts, you still have I/O circuits in the integrated circuit that may be required to interoperate with other ICs at 3.3 volts or 2.5 volts," O'Neill adds.

Getting dosed up

While the destruction of a single satellite can be enormously expensive, far more concerning is the risk of total dose effects, which would impact all satellites with the same design at the same rate.

"What happens with CMOS [semiconductors] during long duration exposure to radiation is that the voltage thresholds of the transistors will shift," O'Neill says. "And so the transistors will become slower, and they'll also become leakier, consuming more power. With modern CMOS this tends to be less of a concern than with older CMOS parts."

That means that, even for power management chips or other components that are not compute-intensive, more modern hardware is preferable. "If all the satellites have this problem, and they all fail at the same rate, then there's your common failure," says O'Neill.

"Now that isn't happening, at least as far as we know, to any of the constellations that have been deployed so far. People are obviously being very careful, but you gotta go into it with your eyes open and just make sure that you're not designing with parts with a total dose problem."

This radiation challenge is "why you can't just fly any old thing in space," he says. "A computing solution may be working very well on the ground, but will it survive the radiation effects? That's not a trivial thing to solve."

Even with radiation-hardened hardware, system-level redundancy is a must. Most satellites have three copies of the same circuit doing the same thing at the same time, and then voting on the result - self-correcting if one has an error.

Another approach is to have more compute in one system, and do the same workload three times simultaneously - which is what the HPE Spaceborne supercomputer did on the somewhat radiation-protected ISS.

No air conditioning in space

With either method, you still face another challenge - how to cool the chips.

"There's no air conditioning in space," O'Neill reminds. "So cooling fans, you have to deal with conductive cooling," essentially using a radiator to radiate the heat out as infrared.

That means that the ability for a satellite to cool is proportionate to its size. "The bigger the satellite is, the more you can dissipate heat. But the bigger the satellite is, the more expensive it is to build and launch."

The larger satellites are also usually larger for a reason - with optics or antennas, or other requirements that may all require power and produce heat.

While terrestrial data centers have seen chip thermal design points (TDPs) soar, this fundamental thermal cap makes a similar densification challenging.

"Having 800 watts in a single device - that's way beyond the threshold of death for most space customers," O'Neill says. "There are limits to how much power can be concentrated into a single integrated circuit before it just becomes impractical to use it.

"There are physical limitations in the 100- to 200-plus-watt range. I'm not going to say impossible, I'm gonna say impractical because there's always ways to do things, but they come with exponentially rising costs and exponentially rising complexity, and potentially with reliability concerns that need to be addressed as well."

A 5GW space data center?

A new industry of space hopefuls has sprung up with the vision of breaking this limit by deploying truly massive satellites. Starcloud, which has raised tens of millions, hopes to one day deploy a 5GW space data center - which it says would require a 2km sq radiator.

For this, and other large-scale approaches, to prove successful requires launch costs falling yet another order of magnitude.

A whitepaper from Starcloud relies on a $30/kg launch cost to reach its $8.2m/40MW deployment figure, but costs are currently around $1,520/kg on the Falcon Heavy.

SpaceX's Starship hopes to bring that to around $150/kg for single-use rockets, and potentially $10-30/kg once it is both reusable and manufacturing costs have dropped. It will also be able to carry larger satellites up in one piece.

Nine test flights have so far ended in explosions, with a tenth last week failing to make it off the launchpad. Each one has yielded crucial information, but it is still too early to know when it will be servicing customers, nor how much it will cost.

Even as we wait for that generational shift to occur, the satellite sector is in a compute upheaval.

"The big change that I'm noticing is the reconfigurability, and the reprogramability of the compute," O'Neill says. "Historically, when a large amount of compute was being deployed in space, it has been done with large numbers of ASICs."

Taking the example of regenerative communication satellites, those companies have built ASICs to do the communications processing as a fixed-function piece of hardware. "Any change in the communication protocol just cannot be accommodated," he says. "And how things are morphing is that those classes of satellites are now trying to do everything they can to make them configurable in flight, so that they can accommodate changes in communication protocol updates, and communications waveforms, things like that.

"They want to have the computing throughput, but instead of having it fixed function, have it reconfigurable," O'Neill says, conveniently noting that AMD's Versal space-grade SOCs can do just that.

"We're getting towards the era of the software-defined satellites. Now I can reprogram and reconfigure it, play around with it on the ground model, get it right, and deploy the change in space - without changing the satellite."