Can datacentres in orbit solve for AI models’ energy demand?, Pg10

Key Highlights:

• Google Research is exploring launching datacentres into space powered by solar energy to address the growing energy demands of AI models.
• AI datacentres require high bandwidth within the datacentre itself, unlike traditional datacentres that need high bandwidth for external communication.
• Project Suncatcher proposes using choreographed clusters of satellites in orbit, maintaining constant line of sight with the sun for continuous power.
• Challenges include managing solar radiation, thermal management in a vacuum, and the overall economic feasibility compared to terrestrial datacentres.

Detailed Insights:

• Traditional datacentres are driven by content consumption, needing bandwidth equivalent to what they deliver externally, while AI datacentres require high internal bandwidth for distributed workloads.
• Microsoft's Fairwater AI datacentre complexes have petabit-per-second links between facilities, highlighting the need for densely networked architecture in AI datacentres.
• Project Suncatcher aims for satellites within a few kilometers of each other, using technologies like multiplexing to maximize data transmission and power efficiency.
• Google is addressing challenges like the impact of solar radiation on tensor processing units, with tests showing Trillium TPUs are surprisingly radiation-hard.
• Thermal management is a significant challenge, as space datacentres must dissipate heat in a vacuum while being constantly exposed to solar energy.
• The economic feasibility hinges on reducing satellite launch costs and achieving power savings to compete with ground-based datacentres.
• Microsoft's Natick project, which explored underwater datacentres for easier cooling, was ultimately abandoned despite its initial promise.

Scientific/Technical Concepts Involved:

• Graphics Processing Units (GPUs): Specialized electronic circuits designed to rapidly manipulate and alter memory to accelerate the creation of images in a frame buffer intended for output to a display device.
• Tensor Processing Units (TPUs): AI accelerator hardware developed by Google, designed specifically for neural network machine learning.
• Multiplexing: Combining multiple data streams into a single channel for efficient transmission.
• Total Ionizing Dose (TID): A measure of the cumulative effect of ionizing radiation on electronic components.