DMS Core is a userspace daemon that reads NVML at 10 Hz on every host: real power draw, junction temperature, and gating decisions, per device, timestamped. It measures first, and acts only where the data justifies it.
The bottleneck for AI infrastructure is no longer GPUs, or capital. It's energized megawatts — and the supply chain for them is measured in years, not quarters.
A single userspace service per host — no kernel modules, no reboot, no changes to tenant code. In shadow mode it writes nothing, so its decisions can be validated before it's ever given authority.
NVML at 10 Hz — power draw, junction temperature, utilization, per device, keyed by UUID and timestamped.
Records what a power governor would have decided, without touching hardware. The safest way to see what it does on a live cluster.
Projects junction temperature 10 seconds forward and flags it early, so an orchestrator can move a workload before a thermal trip, not after.
The advisory path issues zero hardware commands. It publishes; acting on it is your decision, run by your systems.
Most vendors tell you their software works. We'd rather show you what testing on real hardware actually surfaces, including in our own code.
NVML reported under 5% utilization while the card was drawing up to 545 W. An idle H100 with a live context draws 124 W. Our gating logic keyed on that utilization signal — meaning it would have clock-gated a working GPU on 3,952 of 10,848 samples (36%).
We redesigned the gate to require both utilization and power to indicate idle — power vetoes utilization, never the reverse. Re-tested on the same hardware, false gates dropped to zero, while still catching 23.4% of samples as genuinely idle. Gated cards drew 125–135 W; released cards, 129–721 W. Clean separation, no overlap.
The daemon collects the raw signal today. These build on it — and ship the same way everything here did: validated on real hardware before we claim a number.
Coincident-peak analysis across a full fleet, timestamp-aligned at 10 Hz — to measure how much of a provisioned power envelope is real load versus margin no workload has ever touched.
Measuring the efficiency curve of capping GPU power, per silicon generation — how much throughput a cap actually costs, and where the useful knee sits.
48 hours, read-only, one non-production node. The daemon writes nothing to your hardware. You get your real coincident peak at 10 Hz per-GPU resolution — a number facility metering and rack PDUs can't produce — and every figure labeled measured or modeled.
Request an audit