Google首颗轨道AI芯片测试成功
Google把TPU芯片送上天了，用卫星做AI推理，散热和辐射防护都是大挑战。
Google在太空成功测试了首颗轨道AI芯片Trillium TPU v6e，搭载于SpaceX Falcon 9火箭于2026年10月1日发射。该芯片计算能力相当于一个小型数据中心服务器，执行Gemini推理任务，每次运行约15分钟。卫星位于太阳同步低地球轨道，太阳能效率可达地面面板的8倍。
Google’s first orbital AI-chip test is in space and Google has confirmed contact; it is operating as expected.
> launched 1 October 2026 on SpaceX Falcon 9 Transporter-18 from Vandenberg, California.
> Four Trillium TPUs (Google Cloud TPU v6e), roughly the compute of one data-center server / a small TPU slice.
> Workloads are Gemini inference, in bursts of about 15 minutes, then shutdown so the radiators can catch up. Continuous operation is not possible on this prototype.
> Cooling cannot use air or fans. Heat moves through thermal interface material into aluminum/copper heat pipes, then radiators that dump infrared into space.
> Launch loads: the stack sees sustained acceleration around 10g (10 times normal Earth gravity); individual components can see 50–100g.
> Dawn-dusk sun-synchronous low Earth orbit, so the panels see near-continuous sunlight. In that orbit, space solar can deliver up to about 8× the energy of an equivalent ground panel (no atmosphere, weather, or night).
> Google supplied the compute. About refrigerator-sized. Planned life about one year; designed to deorbit within roughly six years.
> the original plan was two purpose-built satellites in 2027. Google put chips on an existing Planet bus to fly this year.
> Radiation: proton-beam tests at UC Davis. Trillium showed no hard failures up to 15 krad(Si). HBM was the weak point, with irregularities after about 2 krad(Si), versus a shielded five-year dose requirement of about 750 rad(Si). Real risk in orbit is single-event bit flips from cosmic rays and solar particles, not just total dose.
> 2027: two satellites to test high-bandwidth laser links between them.
> Long-term design is tight formations, not lone satellites. A research paper models an 81-satellite cluster inside about a 1 km radius, using free-space optical links. A bench demo has already done 800 Gbps each way on one transceiver pair. Close formation is required because optical power falls with distance squared.
> Future craft would carry dozens of TPUs each. Radiators and solar arrays would dwarf the chips — analysts have said radiators could approach airliner scale.
> Scale claim from the program: on the order of 10,000 satellites for a 1 GW orbital cluster. BCG estimates today’s build-and-launch cost for that equivalent at about $30 billion. Earth still wins on cost; Google’s bet is launch-cost decline eventually closes the gap. https://t.co/hro8mrHPlA
