As AI rack densities push past 80 kW and head toward 250 kW, air cooling has run out of room and direct-to-chip (D2C) liquid cooling has become the standard. But a liquid cooling loop is only as good as its weakest hydraulic link — and more often than not, that link is an undersized hose, a flow-throttling manifold, or a coupling chosen without regard to the pressure-drop budget.
Get the sizing right and coolant reaches every cold plate at the flow and temperature the silicon needs. Get it wrong and you face uneven cooling, throttled processors, pump strain, or — worst case — a contaminated loop that clogs cold-plate microchannels across an entire rack.
This guide walks through how to size the fluid-distribution components of a data center liquid cooling loop: hoses, manifolds, and couplings. It’s written for the engineers and procurement teams specifying these systems, with the real numbers the industry uses.
Key takeaway — Sizing starts from flow demand (~1.2 LPM per kW of rack load), then works through hose ID, manifold bore, and coupling size while keeping the total pressure drop inside the CDU’s budget. The two failure modes to avoid: undersizing (throttles flow, strains pumps) and material incompatibility (degrades seals, contaminates the loop). Both are designed out, not discovered in the field.
Step 1: Start From Flow Demand
Everything downstream is sized from one number: how much coolant the rack needs. The industry rule of thumb is a flow basis of roughly 1.2 LPM per kW of rack load.
Worked example:
- An 80 kW rack at 1.2 LPM/kW needs ~96 LPM of total coolant flow
- A 120 kW rack needs ~144 LPM
This flow rate also ties to the thermal target: it’s set so the coolant temperature rise (inlet to outlet) across the server stays around 12 °C. Too little flow and the temperature rise climbs and chips throttle; too much and you waste pump energy and pressure budget.
So the first design step is simple: rack power × 1.2 = your total flow target in LPM. Everything else sizes to deliver that flow without exceeding the pressure-drop budget.
Step 2: Understand the Pressure-Drop Budget
Pressure drop (ΔP) is the constraint that keeps sizing honest. The CDU’s pump can only push so much; every hose, manifold, fitting, and coupling in the loop “spends” some of that budget. Go over budget and you can’t hit the flow target.
Reference numbers the industry uses:
| Component | Typical ΔP guidance |
|---|---|
| Rack-level hoses + fittings | ~34 kPa (~5 psi), conservative |
| CDU secondary-loop max (external) | ~1 bar (~14.5 psi) |
| Cold plate (e.g. NVIDIA H200 class) | <18 kPa (~2.6 psi) at ~8 LPM |
The design job is to keep the sum of all the segment pressure drops — hoses, manifold, couplings, cold plates — within the CDU’s available head at the target flow. The manifold and hoses should contribute as little as reasonably possible so the budget is spent where it does work (the cold plate), not fighting undersized plumbing.
Step 3: Size the Hoses (ID and Dash Size)
Hose sizing is a flow-velocity problem. As the fluid power industry puts it, the hose inside diameter must be sized accurately to obtain the proper flow velocity: too slow and performance is sluggish, too fast and you get excessive pressure drop, system damage, and leaks.
Dash sizes
The fluid power industry uses dash sizes to identify hose ID — the dash number is the hose’s inside diameter in sixteenths of an inch. So -06 = 6/16″ = 3/8″, -08 = 8/16″ = 1/2″, and so on.
Common data center hose sizes
| Segment | Typical hose | Why |
|---|---|---|
| Server blade / cold plate (“spaghetti”) | ~3.5 mm ID | Maneuverable, small OD to save rack space |
| Server node jumper | 3/8″ ID (industry standard) | Handles ~1 GPM per node without excessive ΔP |
| Header / main feed to racks | ~1″ ID | Distributes coolant to many smaller hoses |
The 3/8″ jumper has become a de facto standard because it comfortably handles the ~1 GPM (≈3.8 LPM) flow many server nodes need without causing excessive pressure drops — a good example of matching ID to flow.
Bend radius matters too
A hose that’s correctly sized but kinked is throttled. Server-blade hoses need tight bend radii (a 3/8″ data center hose may have a minimum bend radius around 1.25″ / 32 mm); header hoses don’t need to bend as tightly. Route to avoid kinks — a kink is an unplanned pressure drop.
Step 4: Size the Manifold
The manifold distributes the CDU’s coolant to each cold plate. Sizing it means two things:
- Bore sized for total flow — the manifold trunk must pass the full rack flow (e.g. 96 LPM for an 80 kW rack) without becoming a restriction.
- Branch and port geometry sized for per-server flow — each branch and its coupling must deliver the per-cold-plate flow.
Good manifold design uses dedicated inlet and outlet paths to minimize pressure drop and balance flow across the racks — so the first cold plate and the last cold plate on the manifold get comparable flow. Poor flow balance means some chips run hotter than others.
This is also where material and cleanliness enter: the manifold is usually 304L/316L stainless steel, and it must be passivated, flushed, and sealed so no weld residue or debris flows downstream into the cold plates. (We cover this in our liquid cooling manifold guide and the stainless passivation guide.)
Step 5: Size the Couplings (UQD)
Quick disconnects let servers connect and disconnect from the loop without tools or spillage. The data center standard is the Universal Quick Disconnect (UQD), defined by the Open Compute Project. Key points for sizing:
- Sizes run UQD02 through UQD08 — pick the size that delivers the branch flow without excessive ΔP. The OCP UQD standard dictates baseline flow-to-size ratios so you stay within the ΔP budget.
- Pressure targets: roughly 100 psi max operating, 300 psi minimum burst in the base spec (some commercial UQDs rate far higher — e.g. 290 psi operating, burst over 1,000 psi).
- Material: typically 316 stainless with EPDM seals, for coolant compatibility and corrosion resistance.
- Dripless / dry-break: limits coolant spillage on connect/disconnect — essential for live maintenance.
- Interchangeable: UQD plug and socket halves are compatible across suppliers for a given size, so you’re not locked to one vendor.
Undersizing the coupling is a classic mistake — a UQD02 where the branch flow needs a UQD04 becomes the restriction that throttles the whole branch.
Step 6: Match Materials to Coolant Chemistry
This is the step that, done wrong, contaminates the loop. The wetted materials — hose inner tube, seals, manifold, couplings — must all be compatible with the coolant.
- Typical coolant: a mix of propylene glycol and deionized water (e.g. PG25 = 25% PG), often with biocides and corrosion inhibitors.
- The polarity rule: choose an elastomer with polarity opposite the fluid. Non-polar polymers like EPDM and butyl strongly resist permeation by polar fluids like water — which is why EPDM is the standard hose-tube and seal material for water/glycol loops.
- Verify, don’t assume: proprietary coolant additives make compatibility hard to predict. The reliable approach is physical testing of the tube/seal material with the actual coolant mix. ASTM, ISO, and SAE standards (often borrowed from automotive) serve as guides.
The failure mode if you get this wrong: incompatible seals degrade, small pieces break off, enter the flow stream, and damage the downstream cold plates. A material mismatch upstream becomes a clogged microchannel downstream.
The Sizing Workflow, Summarized
Putting it together, the sequence is:
- Flow target = rack power (kW) × ~1.2 LPM/kW
- Pressure-drop budget = CDU secondary-loop head (~1 bar external), minus cold-plate ΔP, leaves the budget for hoses + manifold + couplings (rack hoses/fittings ~5 psi guidance)
- Hose ID = sized so velocity is in range and ΔP fits (3/8″ jumper, ~1″ header as starting points)
- Manifold bore + branches = trunk passes total flow, branches deliver per-server flow, dedicated inlet/outlet paths for balance
- Couplings = UQD size (UQD02–08) that delivers branch flow within ΔP
- Materials = buyer-specified manifold, seal and tube materials reviewed against the actual coolant; the responsible cooling-system designer approves the final combination
Miss any step and the loop underperforms. Hit them all and coolant reaches every chip at the flow and temperature the design intends.
How DRAmetal Helps
We fabricate drawing-defined fluid-distribution hardware — in-rack, in-row, and row-distribution manifolds in specified stainless grades — against released interfaces and RFQ requirements:
- Manifold bore and branch geometry built to the released flow target and ΔP-controlled design
- Buyer-selected quick-disconnect port interfaces documented on the released drawing
- Wetted materials quoted to the buyer’s coolant-compatibility specification; final system approval remains with the responsible designer
- Passivation, flushing, drying and packaging quoted when the released cleanliness specification defines them
- Pressure / leak testing defined by the buyer’s fluid, pressure, hold-time and acceptance criteria
Related customer project: AI Rack Liquid-Cooling Manifold: Balanced Branch Flow. Related resources: liquid cooling manifold page · 304 vs 316 stainless guide.
Sizing a liquid cooling loop and need the manifolds reviewed? Send us your flow targets, rack layout, and coolant specification. The initial response target is within 24 working hours; sizing feedback, DFM and quotation timing require system and responsibility review.
Need a Custom Sheet Metal Quote?
Send us your CAD drawings — DRAmetal's initial response target is within 24 working hours. Quotation, DFM and lead-time timing depend on file completeness and engineering scope.
GET A FREE QUOTE →
