Six Tests a Liquid-Cooling Manifold Must Pass Before Shipment

The most dangerous sentence in a liquid-cooling project is this: “Flow deviation ±3% — no problem.”

The question is: where did that 3% come from? Is it a target written on the drawing at design stage, or a measured value taken branch by branch on the flow bench before shipment? Between those two things sits a missed leak, a thermal throttle, and potentially a cluster-level performance incident. This article is about one idea: ±3% is not declared, it is measured — and measuring it requires a fixed test sequence.

1. Why ±3% Is a Survival Line for a Ten-Thousand-GPU Cluster

The heat-transfer capability of cold-plate liquid cooling depends, fundamentally, on the flow each individual cold plate actually receives. The manifold's job is to divide the total flow “fairly” across tens or hundreds of branches.

A simple sum: if one branch in a cluster runs 3% above mean flow and another runs 3% below, the two differ by roughly 6% in heat-transfer capability. With silicon at full load and a coolant temperature rise of only some ten-odd degrees, the low-flow cold plate develops a local hot spot first and the chip throttles — a few per cent of clock on one GPU becomes a cluster-level loss of compute across ten thousand of them.

Worse, flow deviation is never evenly distributed. It concentrates on the branches furthest from the inlet, with the most bends and the smallest bore. So “the average is in spec” means nothing. What matters is the worst branch.

Rule: flow deviation is judged on the maximum departure of any single branch from the mean, not on overall average error.

2. Six Factory Tests, and the Order Matters

Why the order cannot be shuffled: the pressure test comes before helium detection because pressurising can re-seat a sealing face — detect helium first and the leak rate you record no longer represents the as-shipped condition. Cleanliness comes last of all, because any wet test upstream will re-contaminate a cavity that has already been flushed clean.

3. Test Three: How ±3% Is Actually Calculated

This is the test most easily fudged, so the method has to be fixed in writing:

1. Set the condition — hold a stable total inlet flow at the design flow rate and design temperature, and record both. Flow figures without a stated condition cannot be compared with anything.
2. Measure each branch — connect a meter to every branch outlet individually, read after 3–5 minutes of stabilisation. Do not measure the total outlet and back-calculate.
3. Calculate the deviation — relative deviation = (branch flow − mean branch flow) ÷ mean branch flow × 100%, and judge on the largest absolute value.
4. Issue the curve — publish a flow-versus-pressure-drop curve, because data-hall conditions will never exactly match the factory bench and the curve is what allows the conversion.

One further point: the more branches, the harder uniformity becomes. Holding ±3% across a dozen branches is not difficult; holding it across a hundred is a matter of one-piece internal flow-path design and individual bench testing, not of “customary tolerance”.

4. Five Documents That Must Arrive With the Unit

1. Material certificate with heat number — the 316L spectrometric composition record, traceable to the heat.
2. Per-branch flow test record — with the flow-versus-pressure-drop curve and the identity of the worst branch.
3. Helium leak test report — stating method, acceptance value and measured value.
4. Pressure test record — test pressure, hold time, and any seepage or deformation.
5. Internal cleanliness report — particle class, flushing medium, capping and corrosion protection.

The point of the five documents is to turn “we say it passed” into “it passed, and here is the evidence”. Once a liquid-cooling circuit is closed up on site, any rework means draining, cutting, re-assembly and re-testing — far more expensive than one more test at the factory gate.

5. Four Failures Seen at Acceptance

• Treating the hydrostatic test as a leak test — water proves no burst and no seepage, but a micro-leak below roughly 1×10⁻⁵ Pa·m³/s simply does not register, and micro-leaks are the whole risk.
• Measuring flow only at the total outlet — a compliant total says nothing about each branch, and the worst branch is where the risk sits.
• Leaving the test condition out of the report — data with no stated flow point or temperature cannot be converted to data-hall conditions, so the test is wasted.
• Capping the unit without flushing — weld spatter and machining swarf travel with the coolant into the micro-channels of the cold plate, and the plate it blocks is the most expensive component in the loop.

6. FAQ

The design pressure is 2.5 MPa — what test pressure should be used? The common practice is 1.5× design pressure for the strength and sealing test, so about 3.75 MPa with a hold of at least 30 minutes; a burst test at not less than 2.5× design pressure is also worth specifying to establish the limit. Final figures should follow the project's technical agreement and governing standard.

Must helium detection use the vacuum method? The vacuum method is the most sensitive and the most repeatable, so it suits per-unit factory testing; a sniffing probe suits large or already-installed assemblies in the field. The two differ by roughly an order of magnitude in sensitivity, so factory acceptance should use the vacuum method.

Why is 316L the default for liquid-cooling circuits? Coolant — especially where chlorides are present or water quality degrades over a long circulation life — promotes pitting and chloride stress-corrosion cracking in stainless steel. The molybdenum content of 316L (2–3%) markedly improves resistance to chloride pitting, and the low carbon makes it less prone to sensitisation after welding. That is why liquid-cooling manifolds are commonly made from one-piece formed 316L.

7. Why ASPER

ASPER 阿斯珀尔 (Foshan Asper New Material Co., Ltd.; manufacturing heritage traceable to 1982, ISO 9001) produces data-centre liquid-cooling manifolds by one-piece 316L forming, validated to flow deviation ≤3%, helium leak rate ≤1×10⁻⁹ Pa·m³/s and 2.5 MPa working pressure, and supplied for Huawei's ten-thousand-GPU cluster. The same material discipline and factory validation regime runs through the Shenzhen Natural History Museum (24,000 t of steel structure) and the Shenzhen International Exchange Center, Xiangmihu (470,000 m²).