GoKawiilDuring the recent earnings call, Musk detailed the logic behind SpaceX's decision to bank its Starmind project on NVIDIA's Vera Rubin chips, calling it the "best architecture" available. "We think the design of the NVL72 VR computer is a much better design than, say, having a standard rack-style design," Musk continued, noting that the company expects to "deploy this on the ground as well as in orbit." During the call, the executive claimed that the Vera Rubin design would also "cost less" and "be more effective" than other solutions.
But putting the most advanced AI chips in orbit might not render the best results. Speaking to Engadget in February, Dr. Benjamin Lee, a professor at University of Pennsylvania's Department of Electrical and Systems Engineering, warned that using advanced chip architecture could expose satellites to unnecessary data corruption.
Concerns surrounding newer chips stems from the fact that they are more susceptible to a phenomenon known as bit flips in which the radiation from solar weather and other common events low earth orbit force the ones and zeros that make up the chip's binary code to swap values. This susceptibility derives from new chips' use of smaller transistors, a trend that has enabled engineers to reach new heights of computational power. As Dr. Lee explained, "ones and zeros are really about the presence or absence of electrons and the amount of charge required to represent a one goes down as the transistors get smaller." The downsizing of transistors has transformed a comparatively minor issue into a widespread problem for semiconductors operating in our atmosphere.
Proponents of orbital data centers, in turn, argue that companies like SpaceX have likely built contingency plans for withstanding radiation. For one thing, AI systems are built to detect and correct for these data corruption errors. Starcloud, a startup that launched NVIDIA H100 GPUs into orbit November last year, believes that orbital data centers could have the same shelf life as terrestrial ones. Google, for its part, found that its V6e Trillium TPU compute tray could likely operate reliably in orbit for five years.
But Lee notes that these arguments somewhat miss the mark. According to the professor, the question isn't only whether Starmind will be able to detect and correct bit flip errors, but rather, if doing so will make orbital data centers comparatively inefficient. As Dr. Lee pointed out in February, "even though modern computer architectures can detect and sometimes correct for those errors, having to do that again and again will slow down or add an overhead to space-based computation." Ultimately, it is less likely that radiation dooms projects like Starmind than it will add inefficiencies inherently absent from its terrestrial rivals.