Manifold Flow Distribution Specs Say 5%, Test Racks Show 10%

The Open Compute Project’s Open Rack V3 Blind Mate Manifold Specification, dated 5 April 2024, highly recommends that coolant flow differ by no more than 5% between branches, and in the same clause reports that the group’s reference design measured 10% in real testing. Every coolant manifold in a direct liquid cooling (DLC) rack, where cold plates on the processors are fed from one shared supply and return pair, splits a single stream into parallel server loops. A branch that gets less than its share carries the same heat in less coolant and runs warmer. The distance between that 5% and that 10%, and what it costs in kelvin, can be worked out from published documents and a few lines of arithmetic.

Flow distribution in the OCP sheet is a spread, not a plus-minus band

The OCP specification defines manifold flow distribution as the maximum branch flow minus the minimum branch flow, divided by the average flow across all branches. That makes it a spread: one number covering the best-fed and the worst-fed branch together.

A rack meeting 5% by that definition can have every branch inside 2.5% of the mean, if the high and low branches sit symmetrically. It can also have one branch 5% short while every other branch sits at the mean.

Korean targets are written in a different form. KHARN, a Korean HVAC trade journal, reported on 2 August 2026 that a national R&D project led by Shinsung Engineering, under the Ministry of Climate, Energy and Environment and the Korea Institute of Energy Technology Evaluation and Planning, lists manifold flow distribution uniformity within ±5% for both single-phase and two-phase manifolds among its quantitative targets.

The same article describes a ±5% limit on branch flow deviation when an in-row coolant distribution unit, the box that separates the facility loop from the rack loop, feeds six to ten racks. The article does not say what the ±5% is measured against, and the project plan behind it is not public in the article.

Read as each branch within 5% of the mean, a ±5% band allows a spread of up to 10% under the OCP definition. That is the value OCP measured, not the value it recommended.

A manufacturer page adds a third form. Vesfil states that its CFD work keeps branch flow deviation within 10% before a prototype is built, without saying whether that is a spread or a deviation from the mean.

Table 1. Branch flow uniformity figures under their stated conditions% of mean branch flow

* Spread: (max − min) ÷ mean branch flow, OCP definition. ±x% read as deviation from mean converts to a spread of up to 2x%.

** As reported by trade press; project plan not reviewed. Same article: ±5% branch deviation for 6 to 10 racks per in-row CDU.

Source: Open Compute Project, “Open Rack V3 Blind Mate Manifold Specification Rev 1.0” (5 April 2024); KHARN, Shinsung Engineering DLC project report (2 August 2026); Vesfil, “Manifolds Engineered for Data Center Liquid Cooling” (modified 27 May 2026)

Without its definition, a flow uniformity figure cannot be compared with another one, and two sheets that both print 5% can describe racks that differ by a factor of two.

The 5% came from a model that assumed identical branches

The OCP group states the basis for its recommendation in the same clause: computational analysis, assuming uniform flow impedance across all branches, quick connectors included, across the intended flow range. It then lists what breaks that assumption in hardware. Blind mate quick connectors vary from unit to unit, branch pressure drops differ (flow through a chassis is the example given), and the manifold’s inlet and outlet design shapes the result.

The group’s advice follows from the gap. A model under ideal conditions should come out low, because the assembled rack will come out higher.

The test conditions matter for reading the 10%. Flow performance tests used Dowfrost LC25, a 25% propylene glycol mixture, with supply on the right manifold and return on the left as seen from the rear, 1 inch hoses and the OCP Large Quick Connector entrance. Top and bottom inlets gave similar branch-to-branch results.

A separate experimental study of a commercial DLC rack, whose abstract appears in the NSF Public Access Repository under the title “Pressure drop analysis of direct liquid cooled (DLC) rack”, found measured maldistribution among server modules and traced it to the manufacturing process of the micro-channel cold plates. The same abstract puts 56% of the rack’s pressure drop in the server modules. It gives the cause of the maldistribution but not its size.

Taken together, the two sources point the same way. Once servers hang on it, most of the resistance and most of the variation sit in the branches, not in the manifold tube. A manifold maker’s CFD figure describes the tube with idealized branches, while the rack’s spread depends on parts the manifold maker does not supply.

Branch pressure drop sets how much header imbalance a rack can absorb

The OCP specification’s remedy is a minimum branch pressure drop. It gives a scale: manifolds with a reasonably large flow section and a longer distance between the inlet and the closest branch should need a minimum branch drop on the order of 1 to 2 psi. The arithmetic below shows why a floor on branch drop, rather than a ceiling, is what holds the spread down.

Assume every branch follows the same law: pressure drop equals k × (flow raised to the power β), with k identical on every branch. Between any two branches, the flow ratio then equals the pressure ratio raised to the power 1/β. A 5% spread means a flow ratio of 1.05.

The exponent depends on what dominates the branch. A paper by Accelsius engineers on two-phase flow restrictors notes that pressure drop scales linearly with flow (β = 1) when viscous shear dominates, as in a long narrow tube, and quadratically (β = 2) through an orifice, where momentum dominates.

Take the orifice-like case, β = 2. Squaring the 1.05 flow ratio gives 1.1025, so the branches may see a pressure imbalance of 10.25% of the branch drop.

On a 1 psi branch, 0.1025 psi of imbalance between the best-fed and worst-fed branch uses up the whole 5% spread.

Now the viscous case, β = 1. The allowable pressure imbalance equals the flow ratio: 5% of the branch drop, or 0.05 psi on a 1 psi branch. A laminar branch tolerates half the header imbalance of an orifice-dominated branch at the same drop.

The header itself supplies the imbalance. The Intel-authored OCP white paper on rack manifolds lists coolant velocity below 1.5 m/s per an ASHRAE guideline, and the OCP cold plate requirements document gives maximum velocities between 1.5 m/s and 2.1 m/s depending on pipe diameter, to avoid erosion.

Velocity head is 0.5 × ρ × v². Assume ρ = 1,000 kg/m³, a round value for a water-based coolant (a 25% glycol mix is somewhat denser), and v = 1.5 m/s.

The result is 0.5 × 1,000 × 1.5 × 1.5 = 1,125 Pa.

Dividing by 6,895 Pa per psi gives 0.163 psi.

That is more than the orifice allowance on a 1 psi branch and more than three times the laminar one. Not all of that head becomes branch imbalance, because friction along a dividing header partly offsets the static pressure recovered as flow leaves it, and the Accelsius model accordingly carries momentum, friction and gravity terms in each manifold segment. The comparison is an order of magnitude, not a prediction for any particular tube.

The practical reading: a header sized to run near the erosion ceiling will struggle to hold a 5% spread on low-drop branches. That is why OCP pairs its 1 to 2 psi figure with a reasonably large flow section rather than with a velocity number.

How much hotter does the short-changed branch run?

In single-phase cooling the answer is an energy balance. Coolant temperature rise across a branch is ΔT = q ÷ (ṁ × cp), where q is the branch heat load, ṁ the mass flow and cp the specific heat. Hold q and cp fixed and the rise scales with the inverse of flow: branch rise = design rise × (average flow ÷ branch flow).

The Intel-authored OCP white paper ties its manifold operating temperature range to ASHRAE W45, a 2°C approach in the CDU and a coolant rise of about 13°C. That design rise is the figure scaled below, and it is the same quantity that sets a loop’s usable capacity.

A 10% spread puts the poorest branch somewhere between 0.95 and 0.90 of average flow, depending on how the other branches are arranged.

In the symmetric case, 13 ÷ 0.95 gives 13.7 K.

In the one-short-branch case, 13 ÷ 0.90 gives 14.4 K.

The penalty is 0.7 K to 1.4 K of extra rise. That is an upper bound for a branch carrying full design heat, and it covers the coolant only: lower flow also lowers the convective coefficient inside the cold plate, which this calculation does not include.

Now place the rack at the W45 edge. A 45°C facility supply plus the 2°C approach gives 47°C to the rack.

Adding the 13 K design rise gives a 60°C mixed return, and the short branch leaves its cold plate at up to 61.4°C.

The OCP blind mate manifold specification states that the overall solution was developed for a maximum working fluid temperature of 60°C. At the warm end of W45, a 10% spread is enough to push one branch outlet past the temperature the manifold family was developed around, while the mixed return reads exactly on target. Nothing at the CDU would show it, since total flow and total heat are unchanged.

In two-phase cooling the penalty stops being proportional. The Accelsius paper models a rack of 34 sleds in parallel on refrigerant R-1233zd(E), calibrated with measured sled pressure drop, at an inlet flow set for an overall exit quality of 0.7.

Without restrictors and with uniform heating, gravity alone spread exit quality from 0.54 at the bottom sled to 0.88 at the top, against a 0.85 limit the authors chose.

Under a nonuniform heating profile, the worst sled reached a quality of 1.76, which the authors identify as dry-out, and 10 sleds exceeded the limit. Orifice restrictors brought every sled under the limit only at a pressure drop twice that of the sled itself. These are model results calibrated on one sled design, not rack measurements.

That asymmetry is the case for writing one ±5% target over both phases, as the Korean project does. In single phase a starved branch costs about a kelvin; in two-phase it can cost the cold plate, which is the other side of what latent heat buys in two-phase DLC.

A uniformity number needs its definition, its test rig and its branch drop

The documents agree that flow uniformity is required and disagree on what a number means. OCP writes a spread and reports that its reference rack doubled the modeled value. The Korean project, as reported, writes a plus-minus band with no stated basis. A manufacturer page writes a deviation figure from CFD. Each is honest on its own terms, and none can be set beside another without a conversion the reader has to do.

The arithmetic adds a second point. Uniformity is bought with branch pressure drop and low header velocity, both of which cost pump energy or tube size, and it is lost in parts the manifold maker does not supply: quick connectors, chassis plumbing and cold plates with manufacturing scatter. A spread quoted without the minimum branch drop it was achieved at, and without saying whether real servers were on the branches, describes the model more than the rack.

The figure I would give weight to is the measured spread on an assembled rack at a stated branch drop and flow, and on that test OCP’s own 10% is the more useful number in its document. For single-phase racks running well inside their temperature class, that 10% costs about a kelvin and rarely decides anything. Near the top of W45, or in any two-phase loop, it becomes a design limit, and those cases call for qualified engineering review rather than a comparison of sheets.

Rebalancing, glycol and field checks on a rack that tests at 10%

Does a 10% spread mean a rack needs balancing valves?

Not by itself. The single-phase arithmetic above puts the cost at 0.7 K to 1.4 K of extra rise on the poorest branch, which matters mainly when the design sits near a temperature limit. The Intel-authored OCP white paper recommends local flow control devices, such as check valves or solenoid control valves, for end-of-row CDU systems, and notes that an in-rack CDU provides flow control to the servers itself. Whether valves are warranted in a given room is a question for the system designer, not for a spread figure alone.

Does propylene glycol change the percentages?

The temperature penalty ratio does not change, because specific heat cancels when every branch carries the same fluid and heat. The absolute design rise does change: the OCP cold plate requirements document notes that conductivity and specific heat fall as glycol content increases, and that viscosity changes alter pump power. Higher viscosity also lowers the Reynolds number in each branch, pushing branches toward the viscous, linear regime that tolerated half the header imbalance in the arithmetic above. The OCP manifold flow tests were themselves run on Dowfrost LC25, so the 10% already reflects a glycol mixture.

Can the spread be checked on a live rack?

Not from the sensors the OCP documents list. The sensor table in the OCP cold plate requirements document asks for volumetric flow at the CDU and at the system level, not at each branch. The test work cited here used dedicated instruments; the Accelsius sled tests, for example, measured each cold plate’s flow with ultrasonic flow meters. A per-branch check therefore means a commissioning step with its own meters, which is easier to plan when the rack is first filled than after a hot branch shows up.

References

• Open Rack V3 Blind Mate Manifold Specification, Revision 1.0 / Open Compute Project / 5 April 2024 / https://www.opencompute.org/documents/open-rack-v3-blind-mate-manifold-specification-rev-1-0-review-april05-2024-pdf
• White Paper: Guidelines to Rack Manifold Requirements and Qualification (Wondimu, Kung, Zhou, Intel Corporation) / Open Compute Project / date not stated on the document / https://www.opencompute.org/documents/ocp-white-paper-rack-manifold-requirements-and-qualification-v3-pdf
• ACS Liquid Cooling Cold Plate Requirements Document, Revision 1.0 / Open Compute Project / 9 October 2019 / https://www.opencompute.org/documents/ocp-acs-liquid-cooling-cold-plate-requirements-pdf
• Investigation of Flow Restrictors for Rack Level Two-Phase Cooling under Nonuniform Heating (Ozguc, Wang, Narayanan, Bonner) / Accelsius / date not stated on the document / https://accelsius.com/wp-content/uploads/Semi-Therm-Honorable-Mention-Award.pdf
• 신성엔지니어링, 300kW/랙 AI DC용 ‘DLC’ 개발 착수 (Shinsung Engineering begins developing DLC for 300 kW-per-rack AI data centers) / KHARN / 2 August 2026 / https://www.kharn.kr/news/article.html?no=31504
• Pressure drop analysis of direct liquid cooled (DLC) rack, abstract as listed on the record for Experimental and CFD Analysis of a Rack Manifold for High Power Density Liquid-Cooled Rack / NSF Public Access Repository / date not listed / https://par.nsf.gov/biblio/10537453-experimental-cfd-analysis-rack-manifold-high-power-density-liquid-cooled-rack
• Manifolds Engineered for Data Center Liquid Cooling / Vesfil / page modified 27 May 2026 / https://vesfil.com/manifolds-engineered-for-data-center-liquid-cooling/