{"id":4591,"date":"2025-11-08T08:30:44","date_gmt":"2025-11-08T02:00:44","guid":{"rendered":"https:\/\/thetysmyanmar.ohofusion.com\/?p=4591"},"modified":"2025-11-11T13:53:15","modified_gmt":"2025-11-11T07:23:15","slug":"project-suncatcher","status":"publish","type":"post","link":"https:\/\/www.thetysmyanmar.com\/en\/blog\/project-suncatcher\/","title":{"rendered":"Project Suncatcher \u2013 A new AI infrastructure being built in space"},"content":{"rendered":"

Project Suncatcher<\/span><\/b> is Google's newest \"moonshot\" project, a groundbreaking project that aims to put Machine Learning (ML) machines on the ground and run them in space.<\/span><\/p>\n

In this project, Google <\/span>Solar-powered satellites<\/span><\/b> to <\/span>TPUs (Tensor Processing Units)<\/span><\/b> combined with <\/span>free-space optical links<\/span><\/b> It is connected to technology.<\/span>
\n The goal is to create a \u201cspace-based AI supercomputer\u201d that can simultaneously run compute on the same scale as data centers on Earth.<\/span>\u00a0<\/span><\/p>\n<\/div>

\"Project<\/span><\/div>

\n

Project Suncatcher's vision<\/span><\/b><\/h2>\n

AI technology is creating a new generation and achieving many scientific and economic achievements.<\/span>
\n But on Earth, due to space, energy, and the environment, <\/span>AI compute<\/span><\/b> It's getting more difficult to expand.<\/span>\u00a0<\/span><\/p>\n

Project Suncatcher<\/span><\/b> is a very nice way to solve this problem.<\/span>
\n Space could provide solar power all day long and not require batteries, making it a more efficient computing environment.<\/span>\u00a0<\/span><\/p>\n

\u00a0<\/span><\/p>\n

System Architecture \u2013 AI Satellite Clusters<\/span><\/b>\u00a0<\/span><\/h3>\n

Project Suncatcher <\/span>low Earth orbit<\/span><\/b> It aims to build a network of solar-powered satellites in low-orbit space.<\/span>
\n Each satellite contains Google's TPU chips, <\/span>optical link network<\/span><\/b> And will contact each one.<\/span>\u00a0<\/span><\/p>\n\n\n\n\n\n\n\n
Component<\/span><\/b>\u00a0<\/span><\/td>\nFunction<\/span><\/b>\u00a0<\/span><\/td>\n<\/tr>\n
Solar Arrays<\/span>\u00a0<\/span><\/td>\nPermanent energy generation through solar<\/span>\u00a0<\/span><\/td>\n<\/tr>\n
TPU Modules<\/span>\u00a0<\/span><\/td>\nOptimized hardware to run Machine Learning<\/span>\u00a0<\/span><\/td>\n<\/tr>\n
Optical Links (DWDM)<\/span>\u00a0<\/span><\/td>\nHigh-speed, low-latency transport<\/span>\u00a0<\/span><\/td>\n<\/tr>\n
Cluster Formation<\/span>\u00a0<\/span><\/td>\nGrouping satellites just a few meters apart<\/span>\u00a0<\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

This approach results in a modular AI infrastructure that can be easily expanded and has redundancy (backup capability).<\/span>\u00a0<\/span><\/p>\n

\u00a0<\/span><\/p>\n

Suncatcher Facing Engineering Challenges<\/span><\/b>\u00a0<\/span><\/h3>\n
    \n
  1. Inter-Satellite Communication Speed<\/span><\/b>
    \n The amount of bandwidth required to run Machine Learning workloads in a space like a data center is simply enormous.<\/span>
    \n Suncatcher <\/span>Dense Wavelength Division Multiplexing (DWDM)<\/span><\/b> With technology <\/span>1.6 Tbps<\/span><\/b> It has successfully passed data transmission tests and is fully practical with stable formation control technology.<\/span>\u00a0<\/span><\/li>\n
  2. Orbital Dynamics & Cluster Stability<\/span><\/b>
    \n To control satellites from a distance of just a few meters <\/span>Hill-Clohessy-Wiltshire equations<\/span><\/b> And <\/span>JAX-based simulation<\/span><\/b> It has been used, and has proven to be able to comfortably handle an 81-satellite cluster at an altitude of 650 km.<\/span><\/li>\n
  3. Radiation Shielding for TPUs<\/span><\/b>
    \n Because radiation is so intense in space, TPU chips need to be perfectly processed.<\/span>
    \n Google's <\/span>Trillium TPU v6e<\/span><\/b> In a proton beam test, <\/span>15 krad(Si)<\/span><\/b> It has been proven to be radiation-resistant because it was able to withstand the impact.<\/span>\u00a0<\/span><\/li>\n<\/ol>\n

    \u00a0<\/span><\/p>\n

    Economic Viability \u2013 Decreasing Launch Costs<\/span><\/b>\u00a0<\/span><\/h3>\n

    Space-based AI was previously not possible and was not cost-effective due to launch costs.<\/span>
    \n However, Google estimates that it could fall below $200\/kg even before the 2030s.<\/span>
    \n At that point, space-based AI could reach a level where it can be used at the same cost as an Earth data center.<\/span>\u00a0<\/span><\/p>\n

    Future Plans \u2013 2027 Prototype Mission<\/span><\/b>\u00a0<\/span><\/h3>\n

    Google has partnered with Planet Labs <\/span>2027<\/span><\/b> The Prototype Satellite is planned to be launched twice this year.<\/span>
    \n In this mission<\/span>\u00a0<\/span><\/p>\n