A TechCrunch article dated October 1 says a satellite launched that day from California on a SpaceX rocket. It is the first time Google has sent one of its advanced chips into space. Planet Labs built the satellite. The flight is meant to show that a Google tensor processing unit, the chip it sets against Nvidia’s GPUs, can work in space. The article describes that as supplying about a kilowatt of continuous power, cooling the chip, and running a series of models. Travis Beals, who manages Project Suncatcher, said they have tested on the ground, and that no ground test fully replaces the flight.

[1]
Ink: a large arc fills the lower sheet, with one small blank rectangle on its crest.
A large arc fills the lower half. One unmarked rectangle sits on the blue band at the crest, and the sky above is empty. That is one chip at the edge of orbit before a formation of satellites exists. An illustration, not a launch photograph., AI-generated illustration, not a news photograph

Once commissioned, the satellite will run the TPU in 15-minute bursts so it does not strain power and thermal systems. This satellite uses a standard Planet Labs platform. The two companies are also building a demo expected next year: two satellites more specifically made for advanced compute, able to run heavier workloads, and trying to work together over a laser link. The same rocket carried more than 100 payloads, including missions from Satlyt and Cowboy Space. The article treats Google’s line as a long project. The orbital data center it describes is a formation of 81 satellites flying close together and processing in parallel. Beals said bandwidth and latency between TPUs matter for a multi-rack workload, and that they are looking at workloads five years out, because the rockets needed to scale orbital data centers cheaply do not exist yet.

[1]

The same day, Google released a peer-reviewed version of its white paper on orbital data centers, to be published in Joule. The authors stress that it is not an economic feasibility study. From a price-cutting curve of about 20 percent a year since Falcon 1, they say it is reasonable to expect launch prices near $200 per kilogram by 2035. Scaling that path by the payload Falcon 9 has already flown would require Starship to lift 370,000 tons. At 200 metric tons a flight, that is about 1,800 launches over ten years, or 180 a year. The headline first said 1,600 and was corrected to 1,800.

Starship has never flown more than five times in a year. Musk has suggested an hourly flight rate in 2029. That suggestion and the paper’s 180 flights a year are different numbers. Google is also a major investor in SpaceX.

[1]

The updated research treats it as likely that the chips can survive space radiation. Google repeated the particle-accelerator tests after finding that the earlier chip layout shielded more radiation than the chips would see in orbit. Logic errors rose slightly. The company still expects the chips to handle large inference workloads across a satellite’s five-year life. Beals said the error rate is very low for typical inference, about one in a million, and that the same rate was already a problem for a mega-scale training run with many thousands of chips running for months.

[1]

要点

  • On October 1 a Planet Labs satellite launched from California on SpaceX, the first time Google has sent a TPU to space. In use it will run in 15-minute bursts.
  • The orbital data center in the piece is a formation of 81 satellites. Next year’s demo is two satellites built more for compute, linked by laser.
  • The paper, to appear in Joule, is not an economic feasibility study. Under the article’s assumptions, Starship needs about 1,800 flights in ten years to match roughly $200 per kilogram by 2035. The headline was corrected from 1,600 to 1,800.
  • Starship has never flown more than five times in a year. After the radiation retest, logic errors rose slightly. Beals put typical inference errors near one in a million, and said mega-scale training is already a problem at that rate.