Documentation › User Guides › Device Owners › Stage 4 › Power, cooling, and noise

Power, cooling, and noise

A device that runs hot earns less. A device that's too loud doesn't run at all because you turn it off. This chapter is about the physical-world side of running a Compute Node for sustained earnings.

Time: 30 minutes to audit; longer if you change cooling. Prerequisites: Stage 3 chapter 4 (monitoring device health) and Stage 2 chapter 6 (power-budget setting) already familiar.

kWh consumption under load

Sustained-load power draw is what matters, not nameplate wattage. Typical ranges:

Multiply sustained load × hours/day × 30 × your $/kWh to get monthly cost. The Stage 4 chapter 3 worksheet uses these numbers.

Thermal throttling and capability_score

The agent reports a capability_score on every heartbeat. When the OS or firmware throttles your CPU or GPU because of temperature, sustained throughput drops — and the next score reflects that. A lower score loses you the higher-tier jobs.

sustained heat → firmware/OS throttle → lower benchmark
              → lower capability_score → fewer/cheaper jobs
              → less earning

You can verify throttling is happening by watching CPU/GPU frequency under load. If frequency drops well below base clock under sustained Zyra work, you're throttling. [VERIFY: capability_score is recomputed on heartbeat from current hardware telemetry — model has capability_score column; how often it re-runs the benchmark is implementation detail]

Passive vs active cooling

• Passive (fanless, large heatsinks): silent, lower sustained capacity, fine for light/medium tasks. Throttles earlier under sustained load.
• Active (fans): higher sustained throughput, audible. Quality of fans matters more than count. One quiet 140mm beats three loud 80mm.

For sustained Zyra workloads, active cooling almost always wins. The exception is a dedicated location where noise doesn't matter and you can over-spec passive.

Fan noise considerations

A device running 16 hours/day with audible fans is unpleasant in a bedroom or shared living space. Practical options:

1. Relocate. A closet, basement, garage corner. Ventilation matters more than insulation.
2. Undervolt. Slight voltage reduction often drops 10-20°C with negligible throughput loss. Worth the BIOS investment.
3. Quieter fans. Noctua / be quiet! premium fans run 15-25 dB lower than stock at the same airflow.
4. Use the power-budget setting. Cap the agent's resource use (Stage 2 chapter 6) so the device doesn't ramp fans to max.

Sustained operation vs occasional bursts

A device built for occasional gaming bursts is not a device built for 16-hour sustained compute. Common surprises:

• Thermal paste from years of normal use is fine for bursts, throttles under sustained. Re-paste if the device is >2 years old and runs hot.
• GPU memory junction temp is the real limiter on modern GPUs. Watch it, not core temp.
• Power supplies sized for peak gaming load are inefficient at the lower-but-constant sustained draw. Efficiency suffers, heat increases.

Dedicated location vs office corner

If you cross 3-5 devices and noise becomes a real cost, consider a small dedicated location: spare bedroom, finished basement, dedicated rack in the garage. Benefits: noise contained, ventilation optimized, cabling tidy, easy physical maintenance. Cost: setup time, possibly added cooling capacity.

For 1-2 devices, an office corner with airflow is fine.

Troubleshooting

• Device stays loud at idle. Fan curve issue — adjust in BIOS or vendor tool. Not a Zyra setting.
• Capability score keeps dropping. Check thermal headroom first; dust buildup is the usual cause.
• Room temperature climbs noticeably. A 300W sustained device dissipates roughly 1,000 BTU/hr. Plan ventilation.

What's next

5. Community and support →

Last reviewed: 2026-05-21