Google’s prototype of its orbital compute satellite took off at present onboard a SpaceX rocket launched from California — the primary time the tech large has despatched one in all its superior chips into area.
Constructed by Planet Labs, the satellite tv for pc will show {that a} Google Tensor Processing Unit, its competitor to Nvidia’s GPUs, can operate in area. Meaning supplying a kilowatt of steady energy, cooling the chip, and working a sequence of fashions by their paces to see if something goes mistaken.
“We’ve achieved testing on the bottom, however you recognize, there’s no check that’s fully pretty much as good as the true factor,” stated Travis Beals, the Google govt managing Project Suncatcher, the tech large’s plan to develop large-scale compute clusters in orbit across the Earth.
As soon as commissioned, the satellite tv for pc will hearth up its TPU in 15-minute bursts to keep away from straining the satellite tv for pc’s energy and thermal administration methods. This satellite tv for pc relies on a normal platform constructed by Planet Labs, however the two corporations are engaged on a demo anticipated to take flight subsequent 12 months that may see two satellites extra purpose-built for superior compute that may run extra substantial workloads. These future variations will try to collaborate through a laser communications hyperlink.
Suncatcher isn’t the one area AI payload on this SpaceX rocket, which is launching greater than 100 totally different payloads, together with missions from Satlyt and Cowboy Area Firm.
What units the Google initiative aside from these startups (and certainly from SpaceX itself) is that it’s a long-term venture.
The main target of this “long-term moonshot,” as Beals places it, is on constructing for the area infrastructure and AI workloads that may exist sooner or later. The corporate envisions an orbital knowledge heart that may be a community of 81 satellites flying in shut formation, processing in parallel.
“The bandwidth and the latency between TPUs actually, actually issues once you’re making an attempt to run a multi-rack workload…we’re making an attempt to stay up for not simply what workloads exist at present, however the place they are going to be in 5 years,” Beals stated. That’s largely as a result of the rockets required to scale up orbital knowledge facilities in an economical manner don’t but exist.
On Thursday, Google additionally launched a peer-reviewed model of its white paper on orbital knowledge facilities, one of the crucial rigorous analyses accessible of how compute will get to orbit. The paper will likely be revealed in Joule.
One of many paper’s most notable elements is how Google thinks about entry to area. Though the researchers stress their evaluation isn’t an financial feasibility examine, it presents an attention-grabbing image of how the corporate sees rockets changing into cheaper over time.
Like all knowledge heart corporations, Google is trying to SpaceX to get its spacecraft off the bottom. (Google can also be a significant investor in SpaceX.)
Arguing that Elon Musk’s rocket builders have achieved a price-reducing “studying curve” of about 20% a 12 months since they launched the Falcon 1 rocket, the authors consider it’s cheap to anticipate the corporate to ship launch costs near $200 per kilogram by 2035.
What is going to it take to do this? Based mostly on the quantity of payload launched by the Falcon 9, they suppose an identical cost-reduction trajectory would require Starship to fly 370,000 tons of payload into orbit. That’s one thing that might take it about 1,800 launches over the following 10 years, or 180 a 12 months — and that’s if it could possibly fly 200 metric tons on every mission.
That’s an enormous ask for a automobile that has by no means flown greater than 5 instances in a 12 months. SpaceX predicts the corporate will likely be flying way over that — Elon Musk has advised Starship may obtain an hourly flight price in 2029, for instance, however Musk says lots of issues.
The excellent news, at the least, in Google’s up to date analysis, is that it appears seemingly that its chips will survive the radiation of area. The corporate needed to redo assessments blasting the chips in a particle accelerator after they realized the configuration of the chips supplied extra shielding than they might truly expertise. This produced barely extra errors within the chip’s logic circuitry, however the firm remains to be assured its chips can deal with massive inference workloads in orbit for the five-year lifespan of a satellite tv for pc.
“The error price may be very low should you’re interested by typical inference operations, proper? Like one in one million,” Beals stated. “However, it was already problematic for doing, say, some mega-scale coaching run the place you’re going to have many 1000’s of chips working for months.”
Correction: The headline for this story initially misstated the training curve estimate for Starship launches as 1,600; it’s 1,800.
Once you buy by hyperlinks in our articles, we could earn a small fee. This doesn’t have an effect on our editorial independence.