Heat rejection caps Google's first orbital TPU satellite at 15-minute compute bursts

The cold of deep space, about minus 270 degrees Celsius, barely helps ^[6]. In a vacuum there is no air or water to carry heat off a chip. Heat leaves only as infrared light from a radiator ^[5]. The Stefan-Boltzmann law, P = e*sigma*T^4, sets the rate, and a perfect radiator at about 300 K sheds roughly 460 watts per square metre per side ^[5]. The limit depends on the radiator's own temperature ^[6].

The International Space Station is the working reference. NASA says its external active thermal control system is "designed to provide 35 kW of heat rejection per loop for a total capability of 70 kW" ^[7]. It uses two rotating wings of three panels, each panel 23.3 by 3.4 metres ^[8]. Together the six panels have about 475 square metres of face ^[1]. The station rejects about 147 watts per square metre of it, roughly a third of the ideal one-side figure ^[2].

Scale up to a ground pod. Google's Ironwood scales to 9,216 liquid-cooled chips spanning nearly 10 MW ^[9]. A dev.to analysis of the launch puts that at about 140 space stations' worth of radiators ^[10]. Its script gives roughly 21,800 square metres of perfect radiator at 300 K, and it ignores absorbed sunlight and Earth's infrared ^[10]. At the station's achieved density, the pod would need about 68,000 square metres ^[3]. Hotter radiators shrink that fast, because output grows with the fourth power of temperature ^[11].

MVP's budget is small enough to check by hand. Shedding a full kilowatt continuously takes about 2.2 square metres of ideal radiator at 300 K, or about 6.8 square metres at ISS density ^[4]. Ars Technica, citing The New York Times, reports that the cooling runs in "brief spurts of about 15 minutes" before the TPUs shut down ^[2]. A run that has to stop so the radiators can catch up means the chips make heat faster than the radiator sheds it while they compute ^[2]. The published accounts do not give MVP's radiator area, the TPUs' power draw, or the idle time between bursts.

The satellite's name, MVP, is accurate ^[4]. Its spacecraft was one Planet had already built for an earlier test ^[14]. The 15-minute figure therefore describes how that bus handles four TPUs. For the number to transfer to the custom satellites, their radiator area per watt of TPU draw would have to match MVP's. I'd expect a bus designed around the chips to run longer. On this evidence, cooling is what limits MVP. That evidence is one secondhand report about one borrowed spacecraft ^[2].

Power is the easy side. Google says a panel in the right orbit sees near-constant sunlight and makes up to eight times more power than on Earth ^[16]. Its launch post calls cooling orbital data centers "a crucial research challenge" ^[1]. Travis Beals, who announced the launch, wrote that Google is "launching a prototype satellite to evaluate how Google Tensor Processing Units (TPUs) perform in space" ^[13].