The rack was left open deliberately. I needed direct access to the Sparks, power supplies, cabling, and fabric while I was still changing the system.
The middle and lower rack initially concentrated the PSU bricks and cable loops on the rack surfaces. The bricks ran warm, the pile restricted the open path through the center of the rack, and the arrangement created the heat pocket I wanted to remove.
The mechanical goal was straightforward: lift every PSU off the shelf, hold it securely against the rack structure, keep both cable ends serviceable, and leave open space below the bricks so the fan could move room air upward.
The orange print proved that the PSU body, cable bend, and connector clearance fit the model.
The PSU fit was right, but the rack-retention clip was not. The first clip did not seat correctly on the shelf lip. That was a local dimensional problem, not a reason to print the entire support repeatedly.
I isolated the clip geometry and printed four small test pieces with four measurements. That let me test the actual rack fit quickly, choose the correct dimension, and move that result back into the full model.
The full print required support material around the cable opening, PSU basket, and rack-retention structure. The orange print documents that support strategy before cleanup.
The corrected orange prototype then seated on the rack and held the PSU with both cable paths accessible.
Once the fit was proven, I printed the production supports in green. Printing them in pairs made the long supported geometry easier to inspect before I removed the supports and installed the parts.
The completed parts mount to the rack, support the PSU body, retain the brick, and leave open volume underneath rather than returning the warm hardware to a flat pile.
The fan sits below the raised PSUs and pushes room air upward through the space opened by the printed supports. The fan remained a separate environmental load; compute and infrastructure power did not inherit a temperature program.
The fan was connected through an AC Infinity eight-outlet environmental power controller. Its rack-local probe supplied temperature, humidity, and VPD. The installed photo records the controller, fan, and probe relationship without publishing the exact outlet selector or controller identity.
The resulting rack is shown from both sides. I reviewed and approved the visible Spark names, power-unit names, equipment model markings, rack labels, and cabling. The public JPEGs retain those pixels while removing the original image metadata.
The test was deliberately narrow. It was meant to prove that the intended fan outlet changed state, that the independent readback agreed, and that the rack probe moved in the expected direction. It was not an airflow calibration.
| Step | Fan state | Rack probe | Interpretation |
|---|---|---|---|
| Stable baseline | ON | 82.4 F |
Starting observation |
| Two-minute bounded interval | OFF | 82.8 F |
+0.4 F with the fan stopped |
| Immediate post-restore read | ON | 82.9 F |
Too early to claim recovery |
| Later independent readback | ON | 82.5 F |
-0.3 F from the OFF reading |
The Raspberry Pi issued only the fan OFF and ON commands. The official app independently verified the relay state and probe value, and every protected outlet remained ON.
Two additional rack/reference snapshots recorded 84.0 F inside versus
78.1 F outside and, later, 82.7 F inside versus 77.9 F outside. Those are
context observations, not part of the matched two-minute differential.
This work demonstrates:
- a measured mechanical fit correction rather than a cosmetic print;
- a repeatable way to support and service all eight PSU bricks;
- clear space below the warm hardware instead of a concentrated pile;
- a dedicated upward fan path;
- identity-bound fan OFF/ON control with independent state verification; and
- a bounded
+0.4 Frise followed by a later-0.3 Frecovery.
It does not claim measured CFM, sealed-rack behavior, long-term thermal capacity, automatic production fan control, or a maximum safe workload temperature.
I followed the bounded fan proof with a longer campaign that captured the rack and all eight Sparks moving from idle into sustained GLM-5.2 inference and back to idle. This run also documented the failures around the experiment instead of editing them out: a misplaced probe, a queued controller read, a later period when both controller snapshots were unavailable, a Spark-only fallback, and an incomplete first version of the graph.
- The top panel shows recorded fleet GPU power in magenta and the number of active GPUs in green. The vertical markers identify first GPU activity, the first all-eight-active sample, and the beginning of cooldown.
- The middle panel shows the fleet mean Spark temperature as the orange line. The pale envelope spans the coolest and hottest Spark in each successful sample.
- The bottom panel shows the two rack/controller probe readings only where a controller snapshot was actually available. The hatched region is missing rack data, not a temperature estimate. The isolated point at the far right is the final rack reading after controller data returned.
The complete record contains 200 records, 198 successful Spark samples, two
failed collection samples, and 50 successful samples with rack temperature.
The other 148 successful Spark samples retain workload and Spark temperature
data but have no rack reading. The hottest Spark reached 86 C; the fleet mean
reached 82.875 C; and observed fleet GPU power peaked at 549.11 W.
This graph proves the recorded workload and Spark thermal progression. It does not prove a numerical fan-cooling effect because verified fan transitions were not present in the campaign samples and rack coverage was incomplete.
The full sequence, including every mistake, safety refusal, recovery action, reporting correction, and remaining uncertainty, is in Rack Thermal Control and GLM-5.2 Campaign: Incident and Recovery.













