Skip to content
Who it is forUniversities and research computing

For universities, research councils and HPC centres

In your words

"The envelope was fixed when the machine was funded, and the queue has outgrown it."

For research computing services, research council programmes and HPC centres: more accelerator-hours inside a funded power envelope, from a runtime measured at up to 21% less GPU die power on NVIDIA H100 NVL.

The situation

A funded envelope, and a community that keeps growing.

The envelope was set by the building, the connection and the funding round, and it holds for the funded period. The machine is sized to it and allocated against it.

Demand from your user community grows faster than the funding cycle. The committee rations, and the next machine is a case to write, score, fund and wait for.

ADAPT is a per-GPU software runtime that changes how much compute the same envelope holds. For a funded service that is counted in accelerator-hours allocated and projects served.

What you are measured on

The lines this decision is judged against.

Allocation
Room for 504 to 760 more accelerators inside the same funded envelope, modelled, so the next capacity bid asks for accelerators rather than power and a connection, and more jobs run for the same community.
Work per watt
On a machine that will not grow this year, the cost line you can still move.
Duty cycle under your scheduler
Gaps between jobs, drain before maintenance windows, reservation edges and quiet periods. A baseline under your own queue measures what they are worth.
A result under your own name
A paired measurement on your machine, with your instruments, that your team publishes on agreed terms.
Governance
Information security and research governance hold the gate on a shared estate, and they see the full install description first.

Released capacity

How it shows up, in megawatts and money.

The instrument's installed fleet quick start, set to a 2 MW machine: 357 NVIDIA HGX H100 8-GPU nodes at 5.6 kW.

From the 48-hour run on NVIDIA H100 NVL

  • measured

    Up to 21%

    Less GPU die power

    NVIDIA H100 NVL, one continuous 48-hour run, managed and baseline arms under an equal power cap, read as NVML GPU die power. Die power is a lower bound on wall power.

  • derived

    Up to +22%

    Tokens per watt

    From die power and serving throughput on the same run: the same work for less energy.

  • derived

    0%

    Throughput change in the run

    The same 48-hour run.

  • observed

    10 to 15 degrees C

    Cooler at the die

    Observed in the same testing.

Modelled on the quick start, whole numbers

  • Power freed inside the funded envelope

    modelled

    About 300 to 420 kW

    2 MW of draw, from the conservative reading to the stated one.

  • Nodes that now fit beyond the machine

    modelled

    63 to 95 nodes

    The power left once managed, in whole nodes of 5.6 kW.

  • Accelerators that now fit

    modelled

    504 to 760 accelerators

    Those nodes at eight accelerators each.

  • Accelerator-hours a year, once added

    modelled

    About 3,090,000 to 4,660,000

    Worked from the quick start's accelerators that now fit, once they are added: 8,760 hours a year at its working utilisation of 70 per cent.

  • Capital not requested

    modelled

    About GBP 3m to GBP 4m, one-off

    Power freed times GBP 9.5m per megawatt, the average project capital cost of a data centre in the Ofgem consultation. Ofgem, Curate: Demand Connections Reform, consultation of 29 July 2026, paragraph 4.12.

A baseline under your own scheduler measures what your queue releases, beside the inference-serving run published here. Each range runs from the conservative reading to the stated one. The conservative reading applies the die figure to the accelerators' share of node draw. The stated reading treats die power as a floor on wall power. Hardware other than NVIDIA H100 NVL stays modelled until a baseline run on it.

Installing it

What installing it involves, and who signs.

For the reviewer: what installs, what it listens on, what it reads and keeps, and how it comes off.

The security review

ADAPT sits under your scheduler and changes what a watt buys, not what runs next. It installs as one service per accelerator under a restricted systemd unit, and stopping the service releases the clocks.

Bring information security and research governance into the first conversation. They receive the unit template, the removal procedure and a software bill of materials before anything installs.

What we bring to a research centre: headroom inside the funded envelope, and a baseline on your machine that your team reports under its own name.

Facility or service director
The funded scope, and the baseline.
Research computing platform lead
The install under the scheduler.
Information security and research governance
What installs, what it touches and how it comes off.
Allocation committee
How released accelerator-hours are allocated.

For the finance lead

  • In a funded programme, released capacity shows up as allocation: more accelerator-hours once the accelerators that now fit are added, inside the same funded envelope.
  • About GBP 3m to GBP 4m of capital not requested, one-off, modelled: a smaller capacity request made earlier, placed as your finance office decides.
  • Every modelled figure here recomputes in the quick start below, on your own nodes, envelope and utilisation.

Case studies that fit

Situations like yours, worked through.

All case studies

Your own numbers

Run it on your site, then talk to us.

The instrument quick start

Installed fleet, set to 2 MW

Opens on 357 H100 nodes in 2 MW. Put in your own nodes, your envelope and your utilisation.

Open this quick start
A conversation

Bring the envelope and the queue. We start with how your machine runs across a week.

Talk about your machine