Assay / Argentum local

GPU-cluster sizing, assayed — power, devices, racks, PDUs & optics for B200 / B300 / GB300 · InfiniBand or RoCE · air or liquid.

Deployment

B200

B300

GB300

InfiniBand

RoCE / Spectrum-X

Air-cooled

Liquid / DLC

GPU count

Compute switch

Frontend oversub (N:1)

Non-GPU fleet

CPU nodes

CPU power kW/node

K8s nodes

K8s power kW/node

Storage

Storage capacity PB

Carrier links

rPDUs / rack (air)

Power / PUE

Air PUE

Liquid PUE (DLC)

Liquid mode applies Liquid PUE to GPU compute and Air PUE to the air-cooled switches / storage / CPU / mgmt.

Use API (exact engine)

Facility power

—

IT power

—

Switches

—

Racks / PDUs

—

Power by category

category	min kW	mean kW	max kW

Switch BOM — by fabric / tier

fabric	tier	model	qty	W each	kW total

Switch capacity reference (ports @ speed)

model	ports (phys/logical)	speed	throughput	breakout

Assumptions

▤ Bill of materials — part numbers for VARs / OEMs

part #	description	qty	conf.

confirmed = OPN on a cited NVIDIA/vendor source · estimate = representative / reach-dependent · unknown = OEM configure-to-order (DGX B300, GB300, HGX). UFM is licensed per managed node/HCA (not per GPU). Validate against a live quote before issuing an RFQ.

◫ Sparing policy — order spares by category

GPU systems %

Network switches %

Fabric manager %

Optics & cables %

Servers %

Storage %

Power / PDU %

item	base qty	spare %	spares

ℹ How this is calculated — formulas & sources