Finding Space. [2/4]

"The universe is under no obligation to make sense to you" - Carl Sagan

Claims that space is excellent for cooling due to the extremely low temperature’s, while intuitively make sense, are only half true (at least for GPU’s as they exist today). Before we get into why, a brief review of the modes of heat transfer. There are three:

The other methods of heat transfer (conduction & convection) rely on matter in order to carry heat. Space is a vacuum (devoid of matter) so we are left with the black sheep of heat transfer…radiation.

An important note from the above equations is that heat transfer for convection and conduction scale linearly (T) while radiation scales exponentially (T⁴). Modern GPU’s like the ones used in AI data centers operate at temperatures between 150ºF and 185ºF. While at much higher temperatures radiation is the dominant mode of heat transfer, at these temperatures conduction and convection are far more prominent.

The temperature of space hovers around -455º F or 2.7º K, which while near the coldest theoretical temperature, doesn’t do us much good since the absence of matter means heat struggles to move into this vacuum.

There is one benefit to this though. It’s cheap from an energy perspective. Using passive radiative coolers and the high temperature gradient in space means we can forego the power hungry chillers and water use typical of data centers.

In the future, if space based computing becomes more mainstream, we could develop specialized GPUs with considerably higher operating temperatures. With high enough operating temperatures radiation could become the more dominant mode of heat transfer and space based cooling could become more advantageous than terrestrial.

One of the more important considerations for throwing these GPUs into orbit will be what the system actually looks like? How will it be organized?

From what I could find, the focus is on mega scale structures with modular GPU containers (I will explain this later). Each mega-satellite must be large enough to function as a standalone data center as data cannot travel fast enough to parallel process across multiple smaller satellites. To see why, we can compare the transfer times using the speed of light in a vacuum v. fiber optic and some basic assumptions around distance.

This 13,000% increase results in massively slower computing speeds and lower utilization of the GPUs if using a constellation of satellites instead of a single mega-satellite (which is equal to data center speeds). It could be possible to make a constellation work although it would require different algorithms than are currently used for parallel processing. It is also worth noting that because of the orbit (600-800km from Earth’s surface) sending data to and from earth will be slower for any space based data center. (Although these speeds are still very feasible)

A space startup called Starcloud has proposed that each of these satellites could be as large as 5GW and use 16 km² of solar panels. They could look something like this:

The solar panel array would function as the primary “spine” with the power, networking, and cooling capabilities. In the center of the array, smaller modules full of GPU racks would dock and undock for maintenance or at end of life. That system could look something like this:

In short, the physical architecture of these data centers is limited by the speed of light which favors mega satellites or extraordinarily tight constellations (< 75m)(likely unfeasible) to achieve comparable processing speeds. This focus on mega scale satellites must be balanced with a need to avoid space debris which I will cover in more depth at the end of the month. If you absolutely can’t wait see this article by Thundersaid Energy for a preview.

Managing weight is undeniably the biggest engineering challenge when sending objects to space. Rockets are designed similarly to soda cans in that a significant portion of their weight is liquid (fuel). With approximately 90% of a rocket's weight being fuel, engineers invest considerable time ensuring the rocket frame is as light as possible to maximize the available payload mass (payload refers to the items the rocket carries that are not part of the rocket or fuel). A great deal of time is also spent optimizing the payload mass to avoid wasting weight. It very expensive to transport even a small payload, so the weight being moved must be essential.

For our data center, this involves using specialized solar panels that are ultra-light, radiation-resistant, and flexible for transport. Another cooling advantage in space is that we don't need to overbuild our cooling system. On Earth, cooling systems must be sized for the hottest potential days. In space, every "day" is the same temperature, allowing our cooling system to be highly optimized, saving weight.

A weight-related downside in space comes from the damaging cosmic rays that our atmosphere typically protects us from. These rays cause GPUs to fail at a higher rate, meaning a greater number of GPUs must be sent to orbit to maintain the same total "amount of compute." (Shielding components can help)

The governing math here is Tsiolkovsky rocket equation and the orbital velocity equation which tell us how much fuel we need in order to reach a desired orbit. For our hypothetical data centers we can make a couple assumptions:

Effectively what this math is doing is 1) establishing the required velocity of our payload to maintain an SSO orbit height of 600 km. If this value is too low our rocket will be pulled to the ground by gravity. We need the centripetal acceleration to exactly match the Earth’s gravitational pull to keep us in a circle. Then 2) using Newtons 2nd and 3rd law to determine how much fuel will be needed to reach to reach said required velocity while accounting for the fact that fuel is spent as the rocket accelerates. As we get higher, the rocket burns fuel and becomes lighter.

If we plot this equation we can see that for every kg of dry mass (weight of rocket without fuel) we would need 11.4 kg of fuel. Making our hypothetical rocket ~92% fuel. This leaves very, very little available mass for our structure and payload.

You might be thinking to yourself, this sounds expensive. You would be right, and I’m super excited to get into that next week. Stay tuned until next Thursday for Economics.