Skip to content
GIRIVON
Infrastructure+
Systems+
Components+
Computing Devices+
Solutions+
Compute Tools+
Knowledge+

AI Infrastructure & Cluster Planner

Turn a target GPU count into a deployable plan: GPU servers, racks constrained by space and power, IT and facility power, cooling, fabric ports, storage and annual energy cost.

From the server's datasheet; for example, NVIDIA reports 10.2 kW maximum for DGX H100.

Results

GPU servers
32
Servers per rack4 by space, 3 by power
3
Racks
11
IT power (max)
326.4 kW
Facility power (max)
457 kW
Cooling
93 TR
Fabric ports (server side)
256
Storage
7,680 TB
Annual energy cost
₹2.24 crore

Power limits each rack to 3 server(s) although 4 would fit physically. Higher-density racks need liquid or rear-door cooling.

What it calculates

  • GPU servers
  • Racks (by space and power)
  • IT and facility power
  • Cooling
  • Fabric ports
  • Storage
  • Annual energy cost

Worked example

Using the default inputs below; change any input above to recalculate.

Inputs
GPUs256
GPUs per server8
Max power per server10.2 kW
Rack units per server8 U
Usable U per rack38 U
Power limit per rack40 kW
Fabric ports per server8
Storage per GPU30 TB
Average utilisation70 %
PUE1.4
Tariff8 ₹/kWh
Results
GPU servers32calculated
Servers per rack3calculated
Racks11calculated
IT power (max)326.4 kWcalculated
Facility power (max)457 kWcalculated
Cooling93 TRcalculated
Fabric ports (server side)256calculated
Storage7,680 TBcalculated
Annual energy cost₹2.24 croreestimated

Methodology

Servers = GPUs ÷ GPUs per server. Each rack holds the smaller of what fits by rack units and by the power limit. Power, cooling, ports and storage scale from the server count; energy uses average utilisation and PUE.

Formulas

Servers per rack
min(⌊U_rack ÷ U_server⌋, ⌊kW_rack ÷ kW_server⌋)

Assumptions

  • Switches and storage are housed in separate racks.

Limitations

  • Network spine racks, storage racks and management infrastructure add space and power.