you are viewing a single comment's thread
view the rest of the comments
[–] 1 point 8 months ago* (last edited 8 months ago) (2 children)

Ya, the economies of how much total space / material for the global network is similar, although lets say higher due to losses in efficiency in distributing it over so many dishes, but in terms of how big any individual radiator is and how much space each one is going to take, the smaller sizes make it easier to manage. Trying to figure out a 150-200m^2^ solar panel radiator is a lot easier than trying to figure out a 1km^2^

The individual power of each satellite having to use a mesh network to train might not be fast enough, maybe they'll still use land based ones for training, but no single person needs more compute than what a satellite can provide. So from the inference / customer computation side of things, it isn't a problem.

edit: I meant radiator, not solar panel

edit: looks like blackwells can run sunstained at 88c, so that will help a bit more as well on size, the calculator now says 103m^2^ instead of 127m^2

  • source
  • parent
  • hideshow 2 child comments
  • [–] 1 point 8 months ago* (1 child)

    So from the inference / customer computation side of things, it isn’t a problem.

    Not necessarily. There are inference schemes where spreading MoE models across 40+ GPUs with a fast interconnect yields better efficiency.

    looks like blackwells can run sunstained at 88c

    The coolant still needs to remain relatively cool to hold that silicon temperature, though. Practically it can't be like 60C.

  • source
  • parent
  • hideshow 1 child comment