Power
Sizing power supplies and racks for GPU systems
From component power to PSU wattage, wall power, rack current per phase and A/B feed design.
GIRIVON Engineering · Updated 19 Nov 2026 · 2 min read
GPU systems draw a lot of power, and short transients can be far above the average. Size from the bottom up.
1. Add up component power
Use rated board power for GPUs and TDP (or package power limits) for CPUs, then add memory, drives, network cards, fans and the platform. Example workstation: two 300 W GPUs, one 350 W CPU and about 150 W for everything else → ≈ 1,100 W sustained.
2. Size the PSU with headroom
Keep sustained load at or below about 80% of PSU capacity and make sure the supply also covers GPU transients: 1,100 W ÷ 0.8 = 1,375 W → a 1,500 W supply. In a server with N+1 redundant supplies, each supply must carry the full load alone.
3. Wall power and efficiency
Wall power = DC load ÷ efficiency. At the 80 PLUS 50%-load requirement for 115 V internal supplies, Gold is 90% and Platinum 92%: 1,100 W DC is about 1,196 W at the wall with a Platinum supply. All of it becomes heat: 1 W = 3.412 BTU/h.
4. Rack current
- Single-phase: I = P ÷ (230 V × PF). A 10 kW rack at PF 0.95 draws 45.8 A — beyond a typical 32 A single-phase feed.
- Three-phase: I = P ÷ (√3 × 400 V × PF). The same 10 kW is 15.2 A per phase.
Continuous loads are commonly kept to 80% of the circuit rating, so 15.2 A needs at least a 19 A rating — a 32 A three-phase PDU is the usual choice.
5. A/B feeds and redundancy
With two independent feeds (A and B), each feed must carry the whole rack if the other fails. N+1, 2N and 2N+1 designs trade installed capacity for resilience; the per-unit load after a failure is the number to check.
Run the numbers in the System Power & PSU Calculator and the Rack Power & Density Calculator. Always follow local electrical codes and a qualified electrical design.