Rack Power Delivery and Busways
A GPU rack has gone from 10 kW to over 120 kW in a few generations, and the electrical design changed with it. At 132 kW on a 415 V three-phase feed a rack draws about 184 amps, which is past what a normal power strip carries, so distribution moves to overhead busway and the rack takes redundant high-current taps. On top of the steady draw sits a synchronized transient every training step, because thousands of GPUs finish a collective at the same instant, and that swing is what sizes the upstream equipment.
TL;DR: Compute the current before anything else, because it decides the distribution design. A rack at 132 kW on 415 V three-phase draws 132,000 divided by the square root of three times 415, which is about 184 A, and redundancy means the facility provisions roughly twice that. That is well past what rack power strips fed from floor whips were built for, so dense deployments move to overhead busway with high-current taps per rack and A and B feeds from separate upstream paths. The second design input is the transient. GPUs in a training job go idle together during a collective and resume together, so a hall of them presents a synchronized load swing at step frequency, and integrators size for peaks in the range of 1.4 times nominal rather than for the average. Power capping through the driver bounds the peak in software when the electrical design cannot.
The arithmetic that sets the design
current from power, three-phase
I = P / (sqrt(3) x V x power factor)
at 132 kW, 415 V, power factor 0.99:
I = 132,000 / (1.732 x 415 x 0.99) = 185 A
what that means for distribution
a common rack PDU tops out well below this, so one strip cannot feed the rack
the answer is either several PDUs per rack or a busway tap rated for the load
and for redundancy the facility provisions A and B feeds each able to carry the whole rack,
so the reserved capacity is about 370 A even though the draw is 185 A
the same rack on other voltages
at 240 V single phase: 132,000 / 240 = 550 A, which is not practical
at 480 V three-phase: 132,000 / (1.732 x 480 x 0.99) = 160 A
sanity: higher voltage is how the current stays manageable, which is why dense halls run
415 V or 480 V distribution rather than the 208 V that older halls were built for
What a rack actually draws, by generation
| Configuration | GPUs per rack | Approximate rack power |
|---|---|---|
| 8x A100 nodes, older hall | 32 to 64 | 10 to 20 kW |
| 4x DGX B300, per NVIDIA's reference architecture | 32 | over 50 kW |
| GB300 NVL72 | 72 | about 120 kW by NVIDIA's figure, 132 to 142 kW nominal with peaks near 155 kW per integrator deployment guides |
The range on the last row is worth carrying around exactly as a range with both attributions, because the two numbers answer different questions. NVIDIA's figure describes the system; the integrator figures describe what a commissioned rack pulls in a real hall with its own networking, storage adapters and margins. Power and Datacenter Constraints covers the facility side of this; the rack-level point is that the difference between 120 and 155 kW is larger than an entire rack of the previous generation.
The transient, and why it is a real problem
why this matters upstream
a normal datacenter load is the sum of many uncorrelated servers, so it averages smooth
a training job is the opposite: every GPU changes state on the same clock
the swing between compute and collective phases can be a large fraction of peak draw
at hall scale this is megawatts moving in milliseconds, which stresses the UPS, the
generators and, at large sites, the grid connection itself
what operators do about it
size for the peak, at roughly 1.4x nominal, rather than for the mean
power-cap the GPUs so the peak is bounded in software: nvidia-smi -pl on a per-GPU limit
or a DCGM policy, at a measurable cost in throughput
add deliberate jitter or a smoothing workload so the fleet does not step in unison
in some designs, on-rack energy storage to absorb the swing
sanity: capping to 90% of board power typically costs a few percent of throughput and cuts
the peak by the same 10%, which is a much cheaper trade than upgrading a substation
Redundancy and how it is actually consumed
- A and B feeds from separate upstream paths, each sized for the full rack, so losing one path does not drop the rack. This doubles the reserved capacity, and reserved capacity is what a colocation contract charges for.
- N+1 at the PDU and busway tap level, so a single tap failure does not take the rack.
- Dual power supplies per node, one on each feed, which is where the redundancy actually terminates.
- The failure mode to plan for is not a feed failing but a feed being taken out for maintenance while the other is at full load, which is the moment a rack discovers it was oversubscribed.
What interviewers are listening for
The current calculation, done out loud, and the transient. Most candidates talk about power in kilowatts and stop; the ones who have deployed talk in amps because that is what the electrician and the colocation provider work in. The transient is the differentiator: an engineer who says "a training hall is a synchronized load, so we size for about 1.4 times nominal and we power-cap when we cannot" has said something that only comes from operating at scale. If cost comes up, the point to make is that redundant provisioning means the contract reserves roughly twice the draw.
Key takeaways
- A 132 kW rack on 415 V three-phase draws about 185 A, and redundant A and B feeds reserve roughly 370 A.
- Higher distribution voltage is how current stays manageable: the same rack is 160 A at 480 V and an impractical 550 A at 240 V single phase.
- Rack power went from 10 to 20 kW on older nodes to over 50 kW for four DGX B300 and roughly 120 to 155 kW for GB300 NVL72, depending on whose figure and what is counted.
- Training loads are synchronized, so size for peaks near 1.4 times nominal rather than for the mean.
- Power capping bounds the peak in software at a few percent of throughput, which is cheaper than upgrading upstream electrical capacity.