Compare the alternatives
ADAPT frees power inside each GPU. Allocation puts it to work across servers.
There are a handful of ways to get more work out of a fixed power envelope. Some trade throughput, capital or downtime; allocation moves power to where the work is. ADAPT works inside each GPU, reading its power signal, and frees power the rest can use.
Side by side
The same envelope, eight different bills.
Eight ways to run a data centre inside the power it already has. Read the last two columns first: they are where the difference lives.
Time to more usable compute
A new grid connection
Years
New electrical equipment
Quarters
More accelerators
Lead time, needs freed power
ADAPT, one software install
Hours
Illustrative, not to scale
Static power capping
- What it does
- One ceiling for every workload, set after the fact.
- Effect on output
- Throughput falls when the cap binds
- Reversible
- Yes
Dynamic power allocation
- What it does
- Moves unused power between servers every few seconds, so more servers fit one envelope.
- Effect on output
- More total output from more servers, each held below full power
- Reversible
- Yes
Workload power profiles
- What it does
- A preset of clocks and power limits chosen for a job, such as Max-Q or Max-P.
- Effect on output
- Max-Q trades a little output for efficiency, and the preset holds for the whole job
- Reversible
- Yes
Scheduler throttling
- What it does
- Defer or slow work to stay inside the envelope.
- Effect on output
- Queues lengthen, latency slips
- Reversible
- Yes
DVFS, the built-in control
- What it does
- Steps between fixed voltage and frequency points after the draw has already changed.
- Effect on output
- Coarse reactive steps leave headroom unused or clip work
- Reversible
- Yes, it is the factory default
Undervolting and clock limits
- What it does
- Hold the part below its rated operating point.
- Effect on output
- Slower, and vendor support gets complicated
- Reversible
- Yes
Liquid cooling retrofit
- What it does
- Remove heat better so density can rise.
- Effect on output
- Unchanged, but capital works and downtime
- Reversible
- No
ADAPT, one software install
- What it does
- Reads each GPU's power signal and sets its operating point many times a second, inside whatever power it is allocated.
- Effect on output
- No throughput change in the run, derived
- Reversible
- Yes, removal restores driver defaults
| Approach | What it does | Effect on output | Reversible |
|---|---|---|---|
| Static power capping | One ceiling for every workload, set after the fact. | Throughput falls when the cap binds | Yes |
| Dynamic power allocation | Moves unused power between servers every few seconds, so more servers fit one envelope. | More total output from more servers, each held below full power | Yes |
| Workload power profiles | A preset of clocks and power limits chosen for a job, such as Max-Q or Max-P. | Max-Q trades a little output for efficiency, and the preset holds for the whole job | Yes |
| Scheduler throttling | Defer or slow work to stay inside the envelope. | Queues lengthen, latency slips | Yes |
| DVFS, the built-in control | Steps between fixed voltage and frequency points after the draw has already changed. | Coarse reactive steps leave headroom unused or clip work | Yes, it is the factory default |
| Undervolting and clock limits | Hold the part below its rated operating point. | Slower, and vendor support gets complicated | Yes |
| Liquid cooling retrofit | Remove heat better so density can rise. | Unchanged, but capital works and downtime | No |
| ADAPT, one software install | Reads each GPU's power signal and sets its operating point many times a second, inside whatever power it is allocated. | No throughput change in the run, derived | Yes, removal restores driver defaults |
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.
Why the difference
Reactive control acts after the excursion. Predictive control acts before it.
A cap can only respond to power it has already seen, so it clips real work. Reading the workload signature first means the draw is shaped ahead of the peak instead of trimmed after it.
The question everyone asks
How is this different from DVFS?
Dynamic voltage and frequency scaling is the control already inside every accelerator. It is a fair challenge, and the difference is fundamental.
DVFS reacts
DVFS watches the draw that has already happened and steps to the nearest point on a fixed table. It is always one excursion behind the workload.
ADAPT predicts
Our method reads the workload signature and shapes the current ahead of the excursion, so the peak never fully forms. Prediction, not reaction.
Steps versus shaping
A fixed table of operating points versus continuous modulation of the draw. One picks from a menu; the other writes the curve.
ADAPT runs alongside the stock driver and firmware. It does not replace DVFS; it removes the conditions that force DVFS to make expensive choices. Removal restores driver defaults.
Next