Space data centres: The next frontier

For many in the data centre industry, space data centres may be a far-off reality. Yet, for AMD, they are the next significant frontier for the industry. Particularly as NASA commits to accelerating its space ambitions, autonomous, high-performance computing becomes mission-critical to enabling the next era of space exploration.

After AMD hardware was included on Orion as part of the Artemis II lunar flyby by NASA (which you can read about on the Capacity website), Capacity spoke with AMD space solutions architect Ken O’Neill about how AMD high-performance and adaptive computing technologies are helping to enable the next era of autonomous space exploration.

“Autonomy is necessary when it comes to space exploration,” O’Neill said. “Communication delays become immense as you move further away from Earth, so you cannot have mission control directing every single space mission – those communication delays alone make that impractical.”

Contextualising the next era of space compute

Communication constraints make automation a ‘must have’ rather than a ‘nice to have’, according to O’Neill. A degree of autonomy is already being developed in this area, having been built into some of the Mars Rover and other ongoing science missions, but O’Neill explained the pace of silicon technologies has caught up with AI.

“I see AI scaling out in multiple different directions in service of autonomy, but it also goes beyond that,” he said. I see it operating at various scales. “An example would be detecting anomalies in spacecraft telemetry data – with this approach, the spacecraft can monitor itself and determine if it needs to take corrective action.”

He also explained how automation can support inferencing solutions, for instance, via cloud detection, or training with space data centres.

“As we proliferate into low Earth orbit, the number of satellites going up is growing dramatically, and with that increase comes a couple of major problems,” he said. “Having AI optimise the use of available RF spectrum is a form of autonomy within the context of how we best use the resources available to our spacecraft or our constellation.”

However, innovation is not without its challenges. While there’s plenty of work ongoing into optical communications in space, there isn’t yet a high-volume, widely deployed optical communications link between satellite and ground, something that O’Neill said is another opportunity for AI-based systems to optimise communications. He added that another big problem with LEO is object detection and avoidance, given that there have already been collisions between satellites in low Earth orbits on account of debris fields.

“Each satellite can be made aware of the historical location and orbit of debris; there needs to be the ability to react in real time,” O’Neill said. “Detection and avoidance is power-intensive and if you’re a satellite, you want to avoid debris without using too much propellant – small course corrections early are much more efficient than large corrections at the last moment.”

He added: “There are programmes to remove debris from orbit, which is another autonomy and AI opportunity.”

Where AMD fits in

As a global semiconductor leader, AMD’s role within this context is to address as much of the space market as possible. As the company evolves and develops new products for the entire space market, O’Neill said its AI-space mission involves putting the most amount of compute capability on orbit as possible. To achieve this, O’Neill explained how AMD needed to increase testing and qualification to enable the existing compute capability to go into more mechanically harsh environments.

“We announced in November that we’re offering a new packaging technology for our product that offers a significant amount of vector processing capability,” he said. “As we scale into deeper space missions, which will be robotic, this technology could lengthen mission duration significantly.”

AMD also launched the Versal AI Edge Generation 2 portfolio, something that O’Neill referred to as “significant upscaling in processing capability”. The company scaled up the number of processors, upgrading the architecture from A72 to A78, and included “safety-critical” automotive enhancements.

“This means you can withstand and survive radiation upsets in space,” he said. “The other enhancement is an evolution of the AI engines integrated into the part – compute capability is increased, but they’re also more power efficient, which is important given that power distribution is difficult.

Another challenge AMD is eager to unpack is thermal. With space being both cold and a vacuum, it is an ideal insulator environment, O’Neill added.

“It’s hard to get rid of heat once it’s being generated on board a spacecraft,” he said. “You’ve got to conduct heat away from the integrated circuits where it’s being generated onto thermal radiators, which then radiate the heat out into space. That’s tricky problem because thermal radiators need to be shaded – in low Earth orbit that’s doubly difficult.”

Making orbital data centres a reality

It’s hard to predict a timeline for space data centres, given how much debate they’ve caused in the market. O’Neill, however, is cautiously optimistic. For him, putting several GPUs into space is not the same as creating a data centre, but rather, the conversation should be measured in megawatts – that while there is opportunity for scale, logistical challenges remain.

“I think we will have data centres in orbit sooner than the pessimists are thinking, but maybe not quite as soon as the optimists are saying,” he added. “There are a lot of difficult engineering challenges that need to be overcome – difficult, but not impossible.”

He added: “There are organisations with funding from venture capital or internal self-funding that are talking about gigawatt-scale. My perspective is that GPUs are probably not the limiting factor, but rather thermal. You don’t just have to power it, you have to cool it, which is challenging.”

O’Neill explained how future data centres will be modular systems launched at different stages and that, to function effectively, communications latency between compute elements needs to be kept short.

“To achieve that, you need autonomous self-assembly of these modules, and at the end of life you need autonomous disassembly as well,” he said. “I think the opportunity for AMD is in the GPU and the compute resources, but also in the communication systems and in the command-and-control systems.”

He added: “I think we’re going to see continued proliferation into low Earth orbit. There is a lot of demand for communications – what we’re seeing with Starlink, we’re seeing that same desire for proliferation across multiple other organisations who also want to put high-volume communication satellites into orbit.”

As the industry sees proliferation beyond low Earth orbit and into medium Earth orbit, O’Neill acknowledged how the space environment becomes more challenging. However, he explained how AMD infrastructure and compute are there to support continued exploration and find new ways to scale in space.

“We provide adaptive SoCs into multiple constellations and individual satellites flying in low Earth orbit and beyond,” he said. “But as proliferation scales up, it’s not scaling into thousands or tens of thousands of satellites – it’s into dozens, or up to perhaps a hundred satellites in a specific constellation. That imposes greater mechanical and radiation challenges, and the components we offer for those missions need to meet them.”

Google targets 2027 for launch of space data centres

Bezos predicts space data centres could solve energy crisis

This feature formed part of the Datacloud Global Congress special edition of Capacity Magazine. Read it HERE.

Datacloud Global Congress is happening now. Find out more below.