Skip to content

Proposal 8: a network of campuses #11

Open

Description

@abu-aska

Closes: the "network" in the objective optifab was opened with. Everything in #2 is one campus.

The gap

optifab models one 10 MW campus and states what a scheduler must respect there: a load floor (about 0.56 MW), a thermal budget (2,500 kWh_th), a margin radius, and a recooling limit (41.5 °C at best). It says nothing about placing work across several campuses.

The system

N campuses, each with its own load floor, thermal budget, margin radius and recooling temperature, and one shared workload. Placing work becomes an optimisation: the most useful compute per unit of utility energy, subject to every campus's gate.

Two things make it more than bin packing:

  • the load floor: a campus that is given too little work stops cooling itself, so spreading load thinly is penalised;
  • the thermal budget: a campus can absorb a burst or ride through an outage, but only until its store is spent.

First slice

Two campuses with different recooling temperatures and one workload with a daily cycle.

Done when

  • the problem is stated as a linear programme and solved exactly;
  • the answer is checked against brute force on a coarse grid;
  • the certificate holds the optimum;
  • a must-fail control shows that the unconstrained optimum breaks a load floor.

Second slice

Dispatch against electricity prices, which funi-frame already lists as open work (aska-physical/funi-frame#2).

Rests on

The contract in optifab. Each campus's gate is the one proved in formal/.

Filed by an AI agent (Claude Code) working for the owner of aska-physical.

Activity

  1. abu-aska commented on Nov 17, 2026

    @abu-aska
    MemberAuthor

    I took a first look at this in #13 before building anything, and it changes what the first slice should be.

    The placement problem as written has no single optimum. In the upstream plant model, utility power is a straight line in IT load: IT + (IT − passive) / COP_loop + auxiliaries + other. Its slope is 1.0357 MW per MW and does not depend on the site. So moving load from one campus to another changes no energy. Any split that keeps every campus above its floor costs the same, and a linear programme over it is degenerate.

    What does differ by site is the load floor. Assumption A-4, made here and not certified: at any recooling temperature the sorption chiller reaches the same share of its reversible COP as at the design point (0.302, from the claimed 0.62 at 58 °C drive, 18 °C chilled water and 30 °C recooling). Then:

    Recooling Sorption COP Load floor Share of nameplate
    30 °C 0.620 0.564 MW 5.6%
    34 °C 0.399 1.127 MW 11.3%
    36 °C 0.325 1.687 MW 16.9%
    38 °C 0.266 2.804 MW 28.0%
    40 °C 0.217 6.113 MW 61.1%
    41 °C 0.196 none: it cannot cool itself at any load

    Two things follow. The 41.5 °C limit in finding 2 is a limit of principle, not a working limit. And a network of one cool campus (30 °C) and one warm one (38 °C) cannot idle below 3.368 MW, which is 16.8% of its nameplate, unless a campus is switched off, and no certificate covers that.

    This is in the certificate as network_first_look, in section 13 of docs/MODEL.md, and guarded by must-fail control 27. The 30 °C itself is the value proposed in aska-physical/funi-frame#4 and is waiting on the owner there.

    So the first slice needs one of these before an optimum means anything. The owners should pick.

    1. A loop COP that depends on the site, from a datasheet or a stated model. Then the slope differs and the linear programme has an answer.
    2. A decision to switch a campus off and on, with its cost. That makes it a mixed-integer problem, and it needs a certified shutdown state that does not exist yet.
    3. A cost for moving work between campuses (latency, bandwidth, state), which needs input from the Weft and Exo owners.

    Option 1 is the smallest. I have not started any of them.


    Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions