The fixed envelope

A research compute facility runs inside an envelope fixed three times over. The building's electrical rating fixes it, so does the import capacity agreed for the site, and so does the funding round that paid for both. A commercial operator with a constrained site has options that cost money. This facility has none inside the programme period. The one thing it can change is what the compute draws.

Software that reduces that draw does not enlarge the envelope. It changes how much compute the envelope holds. For a funded programme that is counted in accelerator-hours allocated and projects served. On a fixed envelope the outcome is growth rather than saving.

First, establish in writing that a service may set accelerator clocks on your estate. ADAPT runs driver-adjacent as one lightweight service per accelerator and needs that authority on the host. On a shared research estate the information-security and research-governance function holds that gate. It will ask for install requirements, a privilege model and a data-protection position. Bring that function into the first conversation, so its answer arrives while the numbers are worked.

Then your load shape. The mechanism is duty-cycle intelligence on bursty fleets. Most of the saving is idle-floor reduction and most of the rest is clock-down at mid utilisation; on a fully saturated fleet it does not transfer. A queue kept full leaves little idle floor, where most of the saving comes from, and the mechanism still works in what remains: gaps between jobs, ramp and drain, maintenance windows, reservation shadows and mid-utilisation phases inside jobs. A three-week validation under your own queue measures how much.

Our coefficient is at the accelerator die and your envelope is bounded at the node input. Between them sits the share of node draw the accelerators account for. On an NVIDIA DGX H100 the published nameplate puts 5,600 W of accelerators inside a 10.2 kW node, about 55 per cent. The runtime does not touch the other 45 per cent.

Two readings follow, and they travel together. The conservative reading applies the die figure to the accelerator share alone and returns about 12 per cent at the node input, modelled. The stated basis treats die power as a floor on wall watts, because a watt not drawn at the die is not converted, distributed or cooled. It returns 21 per cent, modelled. The model carries both readings side by side, and the conservative one is the planning figure.

The programme in this model was funded for 214 nodes and 1,712 accelerators, inside 2,000 kW of IT load at the node input. At nameplate those nodes draw about 2,183 kW, so fitting them needs about 8 per cent off at the node input. Unmanaged, the envelope holds 196 nodes and 1,568 accelerators, 144 accelerators short of the funded scope.

Managed, the same envelope holds 221 nodes and 1,768 accelerators on the conservative reading, modelled. On the stated basis it holds 248 nodes and 1,984 accelerators. The funded scope fits on both, with 56 or 272 accelerators to spare, an uplift of about 13 to 27 per cent. That is node-count arithmetic, and the fit survives at any accelerator share of 40 per cent or above.

A funded programme is measured on hours allocated and projects served. At 90 per cent availability each accelerator gives 7,884 hours a year. The 200 accelerators the conservative reading adds are 1,576,800 accelerator-hours a year; the 416 on the stated basis are 3,279,744, modelled. Over a four-year programme that is 6,307,200 to 13,118,976 hours. At an illustrative median award of 50,000 accelerator-hours, set by your own allocation policy, it is about 31 to 65 further awards a year.

Where your queue is oversubscribed, those hours become further awards, as many as your oversubscription absorbs. Where it is not, the freed envelope banks energy instead, the line below. Most of the conservative gain, 144 of the 200 accelerators, is funded scope the envelope could not hold; the last 56 sit at the edge of the conservative fit and are hardware beyond the funded scope, so plan the funded 214 nodes first.

The freed envelope is spent once. Fill it with accelerators, the largest use where the queue is oversubscribed. Bank it as energy, the smallest use, and often a separately funded line at a public facility. Or hold it as resilience margin. If you fill it, your energy line does not fall and your compute rises. If you bank it, the extra nodes do not appear.

If nodes fit only because the runtime is reducing what they draw, continued operation becomes a standing condition of the configuration. The runtime fails open to full performance on a component fault, the right behaviour at the accelerator. The mitigation lives with the facility: reserved headroom, an agreed curtailment right, or a commitment set far enough below the modelled figure that margin absorbs a total failure.

Energy is computed from your mean operating draw, never from nameplate. At an illustrative 70 per cent of nameplate across the 196 nodes inside the envelope, the avoided energy is 1,413 MWh a year on the conservative reading, modelled. On the stated basis it is 2,574 MWh. At an illustrative 10 pence a kilowatt-hour, your own tariff, those lines are about GBP 141,000 and GBP 257,000 a year. It is the smallest line because a facility on a fixed envelope is short of compute, not electricity.