Google is sending TPUs to orbit, after a five-year radiation dose at UC Davis

Google is sending TPUs to orbit, after a five-year radiation dose at UC Davis

Google is putting Tensor Processing Units into orbit for the first time, on a prototype satellite flying as part of SpaceX's Transporter-18 rideshare mission and built in partnership with Planet. Before that flight, Google ran Trillium TPUs under a proton beam at UC Davis's Crocker Nuclear Laboratory and says they "can survive a radiation total ionizing dose greater than what they would receive during a five-year space mission".

A sun low on the horizon with the text How do we put machine learning in space?

The opening card of Google's Project Suncatcher video series, a sunset over a ridge with the question the project is built around. Image: Google

That five-year figure is a ground-test result, described by Google's September 24 post as "initial results". Nothing has yet been measured in space, which is the entire point of the launch: Google says the mission is "designed to gather in-orbit data on how our TPUs handle the physical stress of spaceflight and the radiation and thermal extremes of space".

What the ground testing covered

The trip to low Earth orbit lasts about 10 minutes, and Google says the spacecraft takes sustained acceleration loads of up to 10 times the force of gravity during it. Individual components, the TPU chips among them, can see 50 to 100 g. The team shook the satellite on all three axes to mimic the frequencies of a launch, and Google says the hardware held up.

Radiation was the second test. The chips ran AI workloads inside the proton beam while the team watched for errors such as a bitflip, which is how the five-year dose comparison was arrived at. Google publishes no dose in rads, and the work has not been peer reviewed.

Cooling is the third problem and the one with no terrestrial answer. "In a vacuum, you can only diffuse heat via radiators, which requires a totally different approach to cooling electronics," Google says. Its current approach combines heat pipes and radiators, tested so far in a thermal vacuum chamber.

Why Google is looking up at all

Project Suncatcher was announced last year as what Google calls "a long-term, research moonshot exploring whether space could one day host scalable machine learning infrastructure". The appeal is power: in low Earth orbit, satellites get near-constant sunlight and can generate up to eight times more solar power than on Earth, at a time when the ground-based build-out of AI compute is being financed on a scale that worries Wall Street.

The satellite going up now carries none of that. Google's designs for later satellites each carry "dozens of TPU chips", flying in clusters and talking to each other by laser, and it puts that link at the precision of "hitting a coin-size target from miles away while both points are in motion". The two-satellite test of the laser interconnect is set for 2027.

What the prototype will not tell us

Google's September 24 post says only that the first TPUs go up "next week", which places the launch in the week beginning September 28. Google publishes no date and no time for it.

It also does not say how many TPUs are aboard the prototype, what orbit it will fly, or what workloads it will run once it is there. The satellite is a rideshare payload, so it goes up when Transporter-18 goes up.

Google calls Suncatcher a research moonshot and this launch a first test, and it does not call any of it a data centre in space. For now the only hardware result on the record came out of a laboratory in California.

Leave a Comment

Your email address will not be published. Required fields are marked *