Autonomous Video Studios
HomePART III · ChannelForge
CHAPTER 16 · CHANNELFORGE

GPU allocation

GPU time is the scarce resource. ChannelForge treats each L4 box as a set of slots and hands them out fairly.

  • Slots. Each L4 box offers two slots; a short takes one, a film takes the whole box for its duration.
  • Scaling. When queued work exceeds free and arriving slots, the fleet starts the cheapest stopped box — cheapest by expected cost per hour, including the risk of pre-emption — with an owner-set preference to break ties or override.
  • Fair share. A free slot goes to paying tiers first, then to the workspace that used the least GPU time in the last 24 hours for its plan's priority, then to the oldest job — so a big workspace cannot starve a small one.
  • Films and shorts. A film earmarks the cheapest box and never pre-empts a running job; while shorts wait, films hold at most 60% of the film-capable boxes, because shorts feed posting slots.
  • Blender. A Blender episode leases a whole box that is set up for Blender, within a share of those boxes and counted against the plan's GPU concurrency, for up to ten hours.
  • Idle stop. A box with no work is stopped after a short grace period; nothing runs that is not paid for.
  • Failure handling. A lost box or a refused start is a capacity error: the job is retried elsewhere, never charged.
  • Quality holds. A rendered episode that fails QA is held and not charged; budgets stop a runaway job with a clear budget error rather than a retry storm.
Figure 31. The admin console's GPU page: boxes, slots, utilisation and costs (development build, demonstration data).