GPU Power Calculator
Estimate the facility power a GPU cluster needs — or how many GPUs a power budget can support — including networking, storage and cooling overhead.
Added to server power. 5–15% is typical.
Cooling and facility losses. 1.1–1.5 for modern sites.
Used for annual energy only.
Facility power
1.87 MW
IT load (servers, network, storage)
1.44 MW
Annual energy at that utilization
13,084 MWh / year
Indicative estimate from published maximum system power, not an engineering design.
Need this capacity? Discuss compute with PermianChainHow it works
Facility power = GPUs × server power per GPU × (1 + networking and storage overhead) × PUE.
- Server power per GPU is the published maximum system power divided by its GPU count: DGX A100 about 6.5 kW, DGX H100 and H200 about 10.2 kW, DGX B200 about 14.3 kW (8 GPUs each), and GB200 NVL72 about 120 kW per 72-GPU rack. It covers CPUs, memory, network cards and fans, not just the GPUs.
- Networking and storage (default 10%) adds switches and storage systems outside the servers.
- PUE (power usage effectiveness, default 1.3) adds cooling and electrical losses. Liquid-cooled modern sites reach 1.1–1.2; older air-cooled sites run 1.5 or higher.
- Annual energy assumes the facility draws that power at the average utilization you enter.
Sources: NVIDIA DGX B200, NVIDIA DGX A100 user guide, and NVIDIA's published DGX H100/H200 and GB200 NVL72 specifications. Figures are maximums; real draw varies with workload.
Frequently asked questions
- How much power does an H100 cluster need?
- With the defaults above (10% networking and storage, PUE 1.3), 1,024 H100 GPUs need about 1.9 MW of facility power, of which about 1.4 MW is IT load.
- How many GPUs can 1 MW support?
- With the same assumptions, 1 MW of facility power supports roughly 548 H100 or H200 GPUs, 391 B200 GPUs, or 860 A100 GPUs.
- What is PUE?
- Power usage effectiveness is total facility power divided by IT power. A PUE of 1.3 means every 1 MW of servers needs another 0.3 MW for cooling and electrical losses.
- Is this an engineering design?
- No. It is an indicative estimate for early planning. A real deployment is sized from the actual hardware, cooling design and site conditions.
Looking for power-ready capacity for a GPU deployment?
Discuss compute capacityEmbed this calculator
Free to use on your own site. Paste this code where the calculator should appear.
<iframe src="https://permianchain.com/embed/gpu-power-calculator/" title="GPU power calculator by PermianChain" width="100%" height="680" style="border:0" loading="lazy"></iframe>