Blind-mate tolerance stack-up.
A blind-mate coupling is asked to close a gap that was created by four different suppliers: the rack fabricator, the manifold machinist, the tray builder and the cold-plate vendor. Published UQDB class parts absorb 0.8–1.3 mm of radial error. This brief shows how quickly a normal rack build consumes that allowance, why statistical and worst-case arithmetic give different answers, where float belongs, and how to stop a misaligned tray from damaging a seal face at service.
DL-008 · Published 2026-09-06 · 9 min read
In a GPU rack the manifold is bolted to structure and the cold plate floats on the board. Every millimetre of assembly tolerance between those two points arrives at the coupling, and the coupling is the only part in the chain that was designed to move. That is the whole idea — but it only works if the budget written on the drawing is smaller than the compensation the part actually delivers.
Most rack programs discover this the expensive way: the first article mates with a firm push, the tenth tray needs a mallet, and by the hundredth unit somebody is replacing scored seals. The arithmetic below is the cheap version of that lesson.
A worked alignment budget.
Seven contributors, each a normal fabrication tolerance. The right-hand column is the same rack after a floating manifold bracket and a datum feature are added.
| Contributor | Bolted build | With float and datum |
|---|---|---|
| Rack frame and rail position | ±0.50 mm | ±0.50 mm |
| Manifold mounting bracket or weldment | ±0.80 mm | ±0.15 mm (floating bracket) |
| Manifold machining and port-to-port position | ±0.35 mm | ±0.35 mm |
| Tray rail and chassis | ±0.90 mm | ±0.30 mm (datum feature) |
| Cold-plate bracket | ±0.50 mm | ±0.50 mm |
| Board and plate flex under load | ±0.30 mm | ±0.30 mm |
| Hose side load on the coupling | ±0.30 mm | ±0.30 mm |
| Worst-case sum | ±3.65 mm | ±2.40 mm |
| Statistical (RSS) stack | ±1.50 mm | ±0.96 mm |
| UQDB class radial compensation | 0.8 – 1.3 mm, published across the class | |
Illustrative tolerances for a bolted 42U-class rack; substitute your own drawing values before using the arithmetic. RSS is the root sum of the squares of the individual tolerances.
Read the two bold rows against the last one. The bolted build lands at ±1.50 mm statistically — already outside the 1.3 mm top of the compensation class, before anyone builds a tray at the edge of two tolerances at once. The worst-case sum, ±3.65 mm, is nearly three times the allowance. After a floating manifold bracket and a datum feature remove the two largest terms, the statistical stack falls to ±0.96 mm and sits comfortably inside the class.
What the budget bought was not a better coupling. Radial compensation is a property of the interface class, so moving from UQDB04 to UQDB08 does not purchase more of it. The fix is mechanical: take the error out upstream, and let the coupling manage the residue.
Radial, axial and angular are separate allowances.
Printing one number on a drawing and calling it "the alignment tolerance" is how a mating panel ends up unbuildable.
Radial — the published number
Published UQDB class parts carry 0.8–1.3 mm of radial compensation: the socket leads, centres the plug and takes up lateral error in any direction. Treat the bottom of that range as the production figure, not the top, and remember that the allowance is consumed by static build error, by hose side load and by thermal movement during operation. Three claimants, one budget.
Axial — the one nobody writes down
The pair has to travel far enough for the valves to open and seal fully. A tray that stops 1 mm short is not "nearly mated", it is a partially mated joint with pressure on one side of a valve that has not seated — the single most dangerous state in the whole interface. The drawing needs an axial engagement travel figure and a pull clearance figure, and both have to be checked with a gloved hand and a hose sweep in the CAD model, inside the 44.45 mm 1U pitch envelope the OCP standard works to.
Angular — ask for it in writing
Angular misalignment between the two axes is the allowance that catalogues rarely publish, and it is the one that damages seals, because an angled entry drags a face across an O-ring instead of seating square. Do not assume a number. Ask the supplier for the angular allowance on the specific part, then control it with the tray rails, parallelism and guide features rather than hoping the coupling absorbs it. If the answer is "the coupling is self-aligning", that is a claim about radial error only.
Where to put float, and where not to.
Float is deliberate, constrained compliance — a bracket that is allowed to move a small distance in a known direction. Used well, it removes the largest term in the stack. Used badly, it adds a new failure mode.
- Float the manifold, not the tray — the manifold is one part per rack; the tray fleet is dozens and will be rebuilt over the life of the loop. Removing ±0.8 mm on the manifold bracket is a one-off change.
- Never let the hose be the alignment system — a compliant hose absorbs position and then quietly applies a side load that consumes the radial allowance and adds to the mate force.
- Constrain the rotation you do not want — a plate free to slide in X and Y but also free to twist will twist. Float in two axes, lock the third.
- One degree of freedom at a time — a bracket that rebounds in X, Y and Z preloads the joint instead of relieving it, and the preload shows up as a higher insertion force.
- Guide first, seal second — a chamfered pin or datum feature should take the mature part of the stack before the seal faces meet, so the coupling's compensation is spent on the last millimetre rather than the first centimetre.
The force budget is where float decisions are paid for. Published UQDB class parts mate at 53–71 N in a single straight push across the size range, with about 62 N at the dash-04 tier. That is a comfortable one-hand figure on a free axis. Add side load from a stack sitting at the edge of its radial compensation and the required push rises — and the person at the rack interprets a joint that will not seat as a joint that needs more force, which is precisely the wrong lesson to teach a service team.
Preventing damage at service.
Couplings rarely fail during operation. They fail during the ninety seconds when a technician is in a hurry.
Damage modes and the countermeasure for each
One more item belongs on the list even though it is not a service action: a fit-check fixture. Before the first production tray is inserted into a live rack, build a gauge that reproduces the worst-case stack and push a coupling through it. Ten minutes with a fixture answers the question that the drawing arithmetic only estimates, and it costs nothing compared with a scrap tray or a scored manifold face.
What to put on the mating-panel drawing.
Six lines. If all six are present, the panel is buildable without a phone call; if one is missing, it will be discovered during first-article assembly.
- Radial allowance — the permitted X and Y offset at the mating plane, stated against the class compensation range.
- Angular allowance — the permitted axis-to-axis angle, confirmed by the coupling supplier for the specific part.
- Axial engagement travel — how far the pair must travel to fully seat, plus the pull clearance to remove it.
- Float specification — direction, total travel and the rotation that is locked out, with the stiffness if a spring or compliant element is used.
- Datum and guide scheme — which feature takes the coarse error and which surfaces are the reference for the port pattern.
- Verification method — the fit-check gauge, the mate force measured at worst-case stack, and the leak or spillage check after the fixture test.
Send those six lines with the port map and the enclosure review becomes a confirmation instead of a correction list — the same principle we apply to every 48-hour DFM review on the coupling side.
Three questions this brief answers most.
How much misalignment can a blind-mate UQDB absorb?
Should I use worst-case or RSS stack-up?
What mate force should a blind-mate tray need?
Send the mating panel. We will check it against the class.
Radial, axial and angular allowances confirmed in writing with the quotation — 48-hour turnaround, MOQ from 50 pieces, fit reports available on sample sets.