Google Contemplates Putting Giant Ai Installations In Low-earth Orbit

Putting AI successful abstraction whitethorn sound for illustration a sci-fi nightmare, but Google is reasoning astir nan thought pinch a investigation endeavor called Project Suncatcher. The thought is to put power-hungry information centers into orbit connected solar-powered satellites, truthful they tin beryllium powered by unlimited, cleanable power disposable 24 hours a day. That would mitigate nan nastiest aspects of AI unreality computing, for illustration nan usage of powerfulness plants that spew immense amounts of CO2 into nan atmosphere.

Project Suncatcher is simply a literal moonshot of nan type that Google used to do much often. The hunt elephantine wants to put its AI chips, called Tensor Processing Units (TPUs), into orbit aboard star panel-equipped satellites. “In nan future, abstraction whitethorn beryllium nan champion spot to standard AI compute,” wrote Google elder head Travis Beals. "In nan correct orbit, a star sheet tin beryllium up to 8 times much productive than connected Earth, and nutrient powerfulness astir continuously, reducing nan request for batteries."

Suffice to say, nan thought poses galore challenges. That proximity to nan sun would expose nan TPUs to precocious levels of radiation that tin quickly degrade physics components. However, Google has tested its existent chips for radiation tolerance and said they'd beryllium capable to past a 5 twelvemonth ngo without suffering imperishable failures.

Another situation is nan high-speed information links of "tens of terabits per second" and debased latency required betwixt satellites. Those speeds would beryllium difficult to execute successful space, arsenic transmitting information astatine agelong distances requires exponentially much powerfulness than connected Earth. To execute that, Google said it whitethorn request to maneuver TPU-equipped satellites into tight formations, perchance wrong "kilometers aliases less" of each other. That would person nan added use of reducing "station keeping" thrust maneuvers needed to support nan satellites successful nan correct position.

The determining factor, though, is money. Launching TPUs into abstraction whitethorn not look cost-efficient, but Google's study shows that doing truthful could beryllium "roughly comparable" to information centers connected Earth (in position of powerfulness efficiency) by astir nan mid-2030s.

While it's presently only a preliminary investigation paper, Google is readying to put Project Suncatcher done immoderate first trials. It has teamed pinch a institution called Planet connected a "learning mission" to motorboat a brace of prototype satellites into orbit by 2027. "This research will trial really our models and TPU hardware run successful abstraction and validate nan usage of optical inter-satellite links for distributed ML [machine learning] tasks," Google wrote.