NodeRAYA watches what your GPUs already report and raises a finding only when a fixed pattern holds. Every finding opens into a config-hashed receipt backed by mathematical algorithms and physics straight from the GPU itself.
No ML agents. No blackbox. No write access. It sits alongside your stack (Grafana, New Relic, Datadog). Fully automated.
Detects GPUs drawing excess power relative to their clock-aware mathematical physics baseline for 10+ minutes.
Reads the GPU's own firmware throttle bitmask directly. Real physics from the silicon, not an inference from room temperature.
Compares identical GPUs in the same fleet using deterministic algorithms. Instantly surfaces the one node behaving differently.
The pattern that fired. The thresholds that were in force. The exact telemetry rows behind it. Nothing is reconstructed after the fact.
Low utilisation sustained through the persistence window while observed power exceeded the clock-aware baseline for this GPU profile.
| Time | Util % | Power W |
|---|---|---|
| 20:00 | 6.4 | 483 |
| 20:05 | 6.6 | 484 |
| 20:10 | 6.9 | 486 |
| 20:15 | 6.7 | 485 |
| 20:20 | 6.8 | 487 |
| 20:25 | 6.5 | 484 |
| 20:30 | 6.9 | 485 |
| 20:35 | 6.6 | 483 |
| 20:40 | 6.7 | 485 |
| 20:45 | 6.8 | 486 |
| 20:50 | 6.5 | 484 |
| 20:55 | 6.9 | 485 |
| 21:00 | 6.7 | 484 |
Utilisation lies outside the same-model cohort distribution by more than three standard deviations. This GPU is behaving differently from its peers.
| GPU | Util % | Role |
|---|---|---|
| NODE 12 | 98.4 | peer |
| NODE 01 | 88.0 | peer |
| NODE 02 | 68.8 | peer |
| NODE 05 | 68.7 | peer |
| NODE 07 | 68.4 | peer |
| NODE 08 | 68.4 | peer |
| NODE 06 | 68.4 | peer |
| NODE 04 | 68.4 | peer |
| NODE 03 | 68.4 | peer |
| NODE 09 | 68.2 | peer |
| NODE 10 | 49.7 | peer |
| NODE 11 | 6.7 | target |
If you run a GPU fleet on Prometheus, we can show you evidence-backed findings on your own hardware within the hour.
contact@noderaya.com