Skip to content

Making sure AI chips don’t overheat in space

Read the Portuguese translation →

Author: Google · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Making sure AI chips don’t overheat in space · Channel: Google · 02:03 · Watch on YouTube

Topics: AI infrastructure and hardware, Robotics and autonomous systems

Project Suncatcher is Google’s effort to put AI compute in space and power it with sunlight. Because space has no air, the project must use radiation to transfer heat from its TPUs and move it from the spacecraft’s interior to radiators. The team is investigating heat pipes and pumped fluid loops, while contending with the large radiator area needed to dissipate the TPUs’ concentrated heat. Before going to space, they will test the integrated payload in a thermal vacuum chamber that simulates a space-like environment.

  • Project Suncatcher aims to put AI compute in space and run it on sunlight.
  • TPUs generate substantial heat in a small area, creating an overheating risk.
  • Cooling systems in terrestrial data centers rely on air, which is absent in space.
  • The spacecraft must use radiation heat transfer to release heat.
  • Heat pipes and pumped fluid loops are being considered to move heat to a radiator.
  • Radiators dissipate a few hundred watts per square meter, while TPUs generate thousands of watts in a hand-sized area.
  • The team expects to need meters of radiator area per TPU.
  • A thermal vacuum chamber will be used to test the payload under space-like thermal and vacuum conditions.
  • The project is an experimental step toward moving data centers off planet and into Earth’s orbit.
Start Chapter Summary
00:00 The cooling challenge Project Suncatcher aims to run AI compute in space, where heat from densely packed TPUs must be safely removed.
00:30 Cooling without air With no air in space, the team must rely on radiation and address the scale of the TPUs’ concentrated heat.
01:01 Moving heat to radiators The team is exploring heat pipes and fluid loops, while accounting for the large radiator area needed per TPU.
01:32 Testing a space-like environment The payload will undergo functional tests in a thermal vacuum chamber as the team works toward its ambitious satellite project.

Project Suncatcher is Google’s moonshot to put AI compute into space and run it off sunlight. — 00:00

Obviously, air doesn’t exist in space, therefore we have to completely change how we cool electronics. — 00:30

We end up having to use meters and meters squared of area per TPU. — 01:01

that’s going to simulate both the thermal and the vacuum environment that we expect to see in space. — 01:32

and important first step towards moving data centers off planet and into Earth’s orbit. — 02:03

Project Sun Catcher, Google, TPU, heat pipes, pump fluid loops, thermal vacuum chamber

00:00 - Project Suncatcher is Google’s moonshot to put AI compute into space and run it off sunlight. (light synth music) One of the challenges we’re trying to solve is how we cool AI chips in space. We have incredibly large amount of power being generated by all these TPUs in a small area. This heat has to make it out safely to space or else we risk overheating the chips. - In a terrestrial data center, we rely heavily on the ability to move air over a hot heat sink.

00:30 Obviously, air doesn’t exist in space, therefore we have to completely change how we cool electronics. - We have to rely on radiation heat transfer, which is a process of which electromagnetic waves are given off of an object. The thermal technology for this project is not new. It has plenty of heritage on prior spacecraft. The challenge here is the scale of things. We’re aiming to have much more power being generated by the TPUs than what’s been done before. So the TPUs live in the guts of the spacecraft. That heat is very concentrated in a small area.

01:01 We have to move it safely from the guts out to the extern of the spacecraft. Right now we’re looking at technologies like heat pipes and pump fluid loops to move it to a radiator. Radiators dissipate on the order of a few 100 watts per meter squared, whereas TPUs generate thousands of watts within something that can fit in your hand. So it’s a very large size discrepancy. We end up having to use meters and meters squared of area per TPU. Before we go to space, we test our solution here on Earth first. - We are going to take it to our thermal vacuum chamber

01:32 that’s going to simulate both the thermal and the vacuum environment that we expect to see in space. We’ll push the integrated payload into the chamber, get it to the right temperature, and start executing functional tests to see how it’s reacting to a space-like environment. - It’s a very ambitious project to make a whole constellation of satellites, but we’re all passionate about the project and are collaborating together to make this happen. - I think everything seems too hard until someone tries it. My first reaction was, “Wow, that seems big even for us.” But it’s an experimental moonshot

02:03 and important first step towards moving data centers off planet and into Earth’s orbit. (soft music)