How do you calculate the power consumption of a GPU rack?

Sum the thermal design power of every accelerator and CPU per node, add about 20% for power supplies, fans, memory and networking, then multiply by the node count: P_rack ≈ nodes × (Σ GPU TDP + Σ CPU TDP) × 1.2. For example, four nodes of 8 × 1 kW GPUs plus 2 × 360 W CPUs is roughly a 42 kW rack — a design ceiling, since real draw follows the workload.

Is TDP the same as actual power draw?

No. TDP is a worst-case thermal rating used to size cooling; typical consumption is lower and workload-dependent. Budget cooling and power on TDP sums (the ceiling), then verify actuals with metered PDUs after deployment.

Formula · 3 min read

TDP → rack kW budget

Rack power is a sum, not a mystery: P_rack = N_nodes × (GPU + CPU + ~20% overhead). A node with 8 accelerators at 1,000 W plus 2 CPUs at 360 W draws ≈ 10.4 kW — four of those is a 42 kW rack before you touch a cooling spec.

GPU accelerator board with copper cold plates over the processors — **PLATE 1** — Where the watts live: a accelerator board with cold plates over the big chips. AI-generated illustration (Z-Image Turbo) — representative, not a specific product.

Formula

P_node ≈ (Σ P_GPU + Σ P_CPU) × 1.2 · P_rack = N × P_node

The 20% overhead covers power supply losses, fans, memory, drives and networking. TDP is a worst-case thermal rating — real draw rides below it and follows the workload, so treat the TDP sum as the design ceiling, not the electric bill.

Worked example

(8 × 1.0 kW + 2 × 0.36 kW) × 1.2 = 10.5 kW/node × 4 nodes = 42 kW/rack

42 kW fits a 50 kW-class liquid cabinet at 25 °C air — but at 35 °C the same cabinet only rejects ≈ 25 kW (derating). The budget and the climate have to be checked together.

Try it

GPU/accelerators per nodeEach (W)CPUs per nodeEach (W)Nodes in rack

10.5 kW/node · rack 41.9 kW IT

⚠️ 42 kW — 50 kW-class cabinet territory, and only in a cool room (derating!)

Common mistakes

Nameplate vs typical. PSU nameplates (e.g., 2 × 3 kW PSUs) state maximum delivery, not draw. Budget on TDP sums; verify on metered PDU data after deployment.
Forgetting the 20%. Fans, memory and converters add up; a rack budget built on chip TDPs alone under-provisions power and cooling by a fifth.
kW ≠ kVA. UPS and electrical work in VA with a power factor below 1; converting late surprises people. Match units early.

Approach temperature (does the climate still fit the budget) · Heat capture split · Rack builder

Sources

Method note: TDP-based ceiling budgeting with 1.2 overhead factor — standard provisioning practice; accelerator classes (700-1,000+ W) per this site's overview. Verify final numbers on metered power.