Power and cooling
Hosting at home
Section titled “Hosting at home”When you run Einsora on a personal machine, or share it (ADR-0014), you pay for its electricity, and your room gets its heat.
| Machine | Typical draw under GPU load | Notes |
|---|---|---|
| Mac mini (Apple Silicon) | Tens of watts | Very efficient; fine to leave on |
| MacBook (Apple Silicon) | Up to the charger’s rating | Shares only on AC power; heavy sustained load warms the battery |
| Mac Studio (Max or Ultra) | Up to roughly 300 W | Check Apple’s published figures for your model |
| Desktop with an RTX 4090 or 5090 | 450–575 W for the GPU, ~600–800 W for the whole system | Loud and hot under sustained load |
monthly kWh ≈ average watts while busy × hours busy per month ÷ 1000For example, a 600 W desktop shared 8 hours a night uses about 600 × 240 ÷ 1000 ≈ 144 kWh a month. Multiply by your electricity rate to see what sharing costs you. Then set caps and schedules in the dashboard to match (power caps are planned).
- Heat and noise: set a schedule (for example, overnight only) for machines in living spaces.
- Laptops: keep them on a hard surface, and let Einsora share only on AC power (the default).
- Power cuts: a small UPS on the control plane machine protects its store.
Datacenter
Section titled “Datacenter”Power is often the real bottleneck, before GPUs or money. Plan for grid instability, high ambient temperatures, and slow utility upgrades.
Rough draw per node
Section titled “Rough draw per node”These are board-power (TDP) figures. Measure your real draw with the BMC under burn-in.
| Config | GPUs | GPU power | Rest of system | Plan for |
|---|---|---|---|---|
| 4× L40S server | 4 × 350 W | 1.4 kW | ~0.8 kW | 2.5–3 kW |
| 8× RTX 4090/5090 server | 8 × 450–575 W | 3.6–4.6 kW | ~1 kW | 5–6 kW |
| 4× A100 80GB PCIe | 4 × 300 W | 1.2 kW | ~0.8 kW | 2.5 kW |
| HGX 8× H100 SXM | 8 × 700 W | 5.6 kW | ~3–4 kW | ~10.2 kW max system |
GPUs spike above their TDP for milliseconds, so PSUs and UPSes need headroom.
UPS sizing
Section titled “UPS sizing”UPS kVA ≥ (sum of planned kW) ÷ power factor (≈0.9) × 1.25 headroomOne 4× L40S node plus a switch and router comes to about 3.3 kW, which needs a UPS of about 4.6 kVA or more.
- Use an online double-conversion UPS for GPU servers.
- Runtime: batteries cover the time until a generator takes over, plus a clean shutdown.
- Generator: needed for anything others depend on. Test it under load monthly.
- Wiring: three-phase supply above about 5 kW, with separate A and B feeds to dual PSUs.
- Monitoring: UPS network cards (SNMP) and BMC power readings, with a graceful drain at a battery threshold.
Cooling
Section titled “Cooling”- Each kW of IT load is about 3,412 BTU/h of heat. A 3 kW node needs roughly 1 ton of cooling by itself.
- Keep server inlet air at or below 27 °C (the ASHRAE recommended range). GPUs throttle and fail early when inlet air is hot.
- Passive datacenter GPUs depend on server fans and front-to-back airflow, so use blanking panels and hot/cold aisle separation.
- Watch GPU temperature and clock-throttle reasons through DCGM.
Colocation versus on-premises
Section titled “Colocation versus on-premises”For a datacenter deployment that others depend on, colocation in a Tier III facility is usually easier than building reliable power and cooling yourself. You get dual utility feeds, a generator, UPS, N+1 cooling, carrier-neutral connectivity, and remote hands. On-premises suits labs and internal setups with modest power needs.
Knowing your running cost
Section titled “Knowing your running cost”monthly kWh ≈ average kW × 24 × 30 × PUE (colocation ~1.5, office ~2+)A node averaging 2.5 kW uses about 2.5 × 720 × 1.5 ≈ 2,700 kWh a month. This is a cost of running the hardware, not something Einsora charges for: Einsora is free.