Testing AI chips to survive in space

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: Testing AI chips to survive in space · Channel: Google · 02:41 · Watch on YouTube

Topics: AI evaluation and benchmarks, AI infrastructure and hardware

Summary

Google’s Project SunCatchers explores using solar-powered AI data centers in space, and tests whether the hardware can survive launch and the space environment. The team uses a vibration table to simulate rocket launch forces and a proton beam facility to test radiation effects while running AI workloads. Tests identified a silent data corruption event, while the TPU performed remarkably well in trials simulating five years in the planned orbit. The project aims to use commercial technology in space without waiting for long design cycles, while accepting the risks of experimentation.

Key points

Project SunCatchers aims to scale AI data centers in space using solar power.
Rocket launch exposes spacecraft to intense vibration and sustained acceleration.
AI chips can experience amplified forces of up to 50 to 100 Gs.
A vibration table shakes hardware packages in three axes to mimic launch.
Cosmic rays and solar events can damage electronics in space.
Radiation testing uses a proton beam while the team runs real AI workloads.
The team detected a silent data corruption event during a beam test.
The TPU held up remarkably well in tests simulating five years in the planned orbit.
The project investigates sending commercial hardware to space without lengthy space-rating design cycles.

Chapters

Start	Chapter	Summary
00:00	Project SunCatchers and spaceflight risks	The team introduces its plan for AI data centers in space and explains the launch forces the hardware must withstand.
00:30	Testing launch vibration	A programmable vibration table simulates rocket launch forces to test whether the custom hardware packaging protects the AI chips.
01:00	Prototype tests and radiation hazards	The prototype is shaken in three axes, then the team describes how radiation in space can damage electronics.
01:33	Proton beam testing and AI workloads	The team explains radiation-related errors and tests hardware with a proton beam while running AI workloads.
02:07	Results and the risks of experimentation	Tests reveal silent data corruption, while the TPU performs well; the team discusses the potential and uncertainty of using commercial hardware in space.

Quotes

Project SunCatchers is Google’s moonshot effort — 00:00

up to 50 to 100 Gs — 00:30

Zero can become a one, one can become a zero — 01:33

The TPU has held up remarkably well. — 02:07

Sometimes these things don’t work out — 02:41

Concepts and entities

Project Sun Catcher, Google, Planet, TPU, single event upsets, total ionization dose, silent data corruption

Full transcript (original language)

00:00 Project SunCatchers is Google’s moonshot effort to scale AI data centers in space utilizing the free solar power available. >> [music] >> One of the very important questions we ask is whether this AI hardware and the system can survive in space. >> It’s a 10-minute ride [music] on a rocket to lower Earth orbit and during that the spacecraft is going to be exposed to [music] tremendous

00:30 vibration forces as well as sustained acceleration loads up to 10 times the force of gravity. Individual components such as [music] the AI chips can actually experience up to 50 to 100 Gs of [music] amplified forces. So really the importance of vibration testing is to ensure that the custom packaging that Planet has designed to prevent this hardware from breaking is sufficient. The main piece of equipment we need is a vibration table that can be programmed with the frequencies [music] that will mimic the rocket launch. So we’ll build

01:00 our first prototype of a package >> [music] >> and then put it on a vibe table and we will shake it in all three axes. These early prototype or or kind of moonshot vibe tests, they rarely go as you expect them to [music] and so I think we were pleasantly surprised how well it went. Once you get into space, now you’re subjected to solar events as well as cosmic rays. Essentially [music] high-energy particles that can wreak havoc on your electronics. >> We are protected in some ways from this radiation because of the Earth’s

01:33 atmosphere. We live underneath this beautiful blanket, right? But in space, it can create severe damage. You can have something like a bit flip. Zero can become a one, one can become a zero and these are actually what are called single [music] event upsets. So that’s one kind of error and these are correctable. The other is total ionization dose >> [music] >> and this is the damage of it being constantly under this radiative environment. >> [music] >> And the way we test for this is we take it to this proton beam facility and bombard it directly with these protons.

02:07 While we are doing this, we actually run actual AI workloads and we see how the errors can affect [music] our training. >> So, our tests have found what is called an SDC or it’s called a silent data corruption. >> [music] >> It is the first time we have found actually an SDC during a beam test. >> So, we were testing for if the TPU would survive 5 years of the kind of orbit [music] that we are looking at and each part has different survival rates and different error rates. And the TPU has held up remarkably well.

02:41 >> The importance here is that we can leverage the advances in the [music] commercial industry and immediately send those to space instead of waiting for long design cycles [music] to make it space rated. >> These are risky projects. Sometimes these things don’t work out, but if you are a scientist, you’re ready to kind of face that challenge and solve the problems that come along the way and you might end up with something completely different or new, which is essentially the fun part of working in something like this.