The words this site uses precisely.

Power envelope

The most power a site, a hall or a rack is allowed to draw.

Set by the grid connection, the switchgear and the contract, and fixed for years. Everything on this site is about fitting more work inside the envelope you already have, not about getting a bigger one.

See also Headroom, Site IT power

Headroom

Power you already pay for that your hardware is not using, made available again.

When each GPU draws less for the same work, the difference is capacity inside the envelope. We call that headroom and hand it back to the operator; it is the quantity the instrument reads.

See also Power envelope, Up to 21%

Stranded power

Power a site has provisioned that its GPUs are not drawing at that moment, because demand swings with the workload.

Power set aside for each rack or node at its peak sits unused whenever the work in hand needs less. Allocating power across nodes as demand moves puts it back to work. ADAPT frees more of it one level down, inside each GPU, by setting the operating point to what the work needs; we call what it frees headroom.

See also Headroom, Power envelope

Site IT power

The power that reaches the computing equipment, before cooling and losses.

The instrument's slider is IT power. Facility power is IT power multiplied by the PUE; the two are kept apart on every page so a die-plane figure is never quietly applied to a facility figure.

See also PUE, GPU die power

PUE

Power usage effectiveness: how much power a facility draws for each unit its computers draw.

A PUE of 1.25 means the site draws 1.25 MW at the meter for every 1 MW the IT equipment uses. The instrument uses your own PUE to turn released IT power into energy and cooling load.

See also Site IT power

GPU die power

The power the GPU chip itself draws, as the card reports it.

Our measured figure is GPU die power read through NVML on the card. It is a lower bound on the power saved at the wall, because the losses between the die and the meter scale with it. We publish the die figure and let your own meter confirm the rest.

See also Wall power, Measured, derived, observed, stated, modelled, designed, projected

Wall power

What the meter sees: the die, the board, the power supplies and the losses between them.

The published figure is GPU die power, a lower bound on the saving at the wall. During the three week validation on your own racks, your own meter reads the wall.

See also GPU die power, Three week validation

Up to 21%

How much less GPU die power the managed GPUs drew than the unmanaged ones, over 48 hours on H100 NVL.

Measured on production NVIDIA H100 NVL with managed and baseline arms under an equal power cap. It is the one measured coefficient on this site, and every other figure is arithmetic on it or your own inputs.

See also Measured, derived, observed, stated, modelled, designed, projected, Paired arms, equal cap

Tokens per watt

How much useful output a fleet produces for each watt it draws.

Derived from the measured die power and the serving throughput of the same run: the same work for less energy, so the ratio rises by up to +22%.

See also Up to 21%, Measured, derived, observed, stated, modelled, designed, projected

Paired arms, equal cap

Two groups of identical GPUs running the same work at the same time, one managed and one not, under the same power limit.

Both arms run at once, on the same hardware, under the same cap. The proof page shows the arms and the window.

See also Measured, derived, observed, stated, modelled, designed, projected, Baseline

Baseline

How your fleet behaves before anything of ours runs on it, recorded first.

The three week validation starts with a baseline on your own racks. The figure on this site becomes your figure only against that record; without it there is nothing to compare.

See also Three week validation, Paired arms, equal cap

Measured, derived, observed, stated, modelled, designed, projected

Seven words that say how much to trust a figure, printed beside every figure on this site.

Measured: read from hardware in a defined test. Derived: arithmetic on measured figures. Observed: seen during the same defined test as a measured figure. Stated: a fact about our programme or product, such as days of testing, not a reading from hardware. Modelled: computed for hardware we have not measured, until a baseline run on it. Designed: a capability specified and not yet built. Projected: expected from a design target or the roadmap, and not yet reached in a test.

See also Up to 21%, GPU die power

Adaptive Current Modulation

The method: read a machine's power signal in the frequency domain, predict where it goes, and set the operating point ahead of it.

Reactive regulation corrects an error after it appears. ACM decomposes the draw into its spectral components and acts on each, so the current follows the work instead of chasing it. The subject of pending patent applications, filed and examined: the UK Intellectual Property Office found claims 5 and 15 novel and inventive.

See also Operating mode, GPU die power

Operating mode

A preset that says how hard the runtime trades power against throughput.

Performance holds throughput first. Balanced is the production default. Green takes the most power out. The presets ship with the runtime and the assessment reads their target bands.

See also Adaptive Current Modulation

Power profile (Max-Q, Max-P)

A preset of GPU clocks and power limits for a kind of work: Max-Q for the most work per watt, Max-P for the most work.

A profile is usually chosen when a job starts and holds while it runs. ADAPT sets each GPU's operating point continuously, from the GPU's own power signal, so the setting follows the work as it changes; its operating modes play the part of the profile's objective.

See also Operating mode, Stranded power

Fails open

If anything goes wrong, the runtime steps aside and the GPU runs as it did before.

Any fault, any removal, any stop returns the driver's own defaults. Nothing of ours has to be running for your fleet to work.

See also Driver adjacent

Driver adjacent

The runtime sits beside the GPU driver, not inside it, and never touches your workloads or data.

It reads the card's own telemetry and sets the card's own operating point through the vendor's public interface. No driver change, no firmware change, no access to models or tenants.

See also Fails open

Neocloud

A cloud provider built around GPUs for AI, selling compute by the hour.

Neoclouds sell megawatts as accelerator hours, so capacity released inside their envelope is revenue. They are the buyers for whom headroom pays back fastest.

See also Headroom

AI factory

A data centre built to run AI models at scale, where power is the binding constraint.

Whether you run one or are building one, the power plane decides how much compute fits. ADAPT reduces what each GPU draws inside it; for a build in planning, we size the plane before the concrete.

See also Power envelope, Site IT power

Three week validation

A baseline on your fleet, a paired evaluation, and a report you keep, in three weeks.

Week one records your fleet as it is. Week two runs the managed and unmanaged arms side by side. Week three writes it up with the records attached. No procurement is needed to start.

See also Baseline, Paired arms, equal cap

These words at work, at greater length: the guides to energy efficiency, PUE, tokens per watt and data centre power capacity.

Measured on production NVIDIA H100 NVL under an equal power cap on managed and baseline arms. Your fleet baseline confirms the figure for your site.