92 Days Under Load
The evidence required to report a dense-compute trial's uptime, utility event, thermal behavior, power measurements, and limits.
- Published
- Filed under
- incident review
The Conture dense-GPU trial ended on June 4, 2024, after 92 consecutive days. Its canonical result is precise: the two DGX A100 systems recorded 100% service uptime during the trial. One 23-second utility-power event occurred on April 17, and the UPS carried the protected load without a service interruption.
Those statements are not interchangeable. Utility availability describes the incoming source. Protected-power continuity describes the facility power path. Reachability is one monitoring signal. Service uptime is the customer-facing outcome under the agreed definition. A publishable report has to keep all four separate.
The availability statement
The final calculation should show the trial window, time zone, total measured minutes, monitoring cadence, maintenance treatment, data gaps, incident criteria, and numerator and denominator used for service uptime. It should identify which signals came from facility monitoring and which came from Conture’s application or system checks.
Helixrack’s monitoring watchdog checked both systems every 30 seconds. That supports a reachability record, while Conture’s agreed acceptance criteria define the 100% result.
The article must not use “99.98%.” That figure conflicts with the resolved record and would also imply downtime the trial did not have. The supported formulation is 100% service uptime, subject to the documented measurement definition.
Twenty-three seconds on protected power
The April 17 event lasted 23 seconds at the utility layer. The UPS absorbed the interval, so neither DGX lost service. The final report should reconstruct the sequence from synchronized timestamps: utility anomaly detected, protected power assumed the load, utility returned or the input stabilized, and normal state resumed.
That chronology must come from logs, not from a dramatic illustration. A generated graphic should show a clean source-state timeline and the two 6U DGX A100 systems at conditioned-air supply with hot exhaust directed toward return. It should not show sparks, shutdowns, later equipment, or customer-visible downtime.
Thermal and power findings
Each DGX A100 had dual 40A, 208V A/B feeds and a published maximum system draw of 6.5kW. The period DGX A100 announcement establishes the 640GB product configuration; internal electrical and metering records must establish what these particular systems drew.
The publishable report should chart measured A- and B-side load, peaks and percentiles, inlet and exhaust temperatures, room conditions, sensor positions, and any alert excursions. It must distinguish nameplate or published maximum from observed demand. It should also state whether testing included feed loss, workload variation, maintenance, or abnormal ambient conditions.
NVIDIA announced the Blackwell platform while this trial was underway. That is period market context, not proof that Blackwell hardware was present, ordered, or supported at Helixrack. No Blackwell equipment belongs in the trial imagery or capacity conclusions.
The threshold, not the trophy
The useful output of a dense-compute trial is a boundary for future intake. The final report should state what combination of rack position, feed capacity, cooling path, monitoring, and operating reserve Helixrack approved—and where it would require a new engineering review.
Ninety-two days under load did not make every open rack unit suitable for dense compute. It established evidence for two specific systems in a specific configuration. That narrower conclusion is more valuable because an operator can apply it to the next intake decision without pretending space alone equals capacity.
Sources
- NVIDIA announces the DGX A100 640GB configuration November 16, 2020 · period
- NVIDIA announces the Blackwell platform March 18, 2024 · period