Tolerance Stack-Up
Tolerance Stack-Up: Why the Parts That Fit on Screen Don't Fit in Metal
Every part is a little off, and the little offs add up. Ten parts at ±0.1 mm stack to ±1.00 mm worst case and ±0.32 mm statistically, and the difference between those two answers is the difference between a product that assembles and a product that doesn't. The fix is rarely a tighter tolerance. It is usually a shorter chain.
First, a simple picture
A staircase, not a step.
One step is easy: a single part, a single size, a single measurement. Five steps are a staircase, and a staircase is only as true as the sum of its steps. If each step is a little too tall or a little too short, the top of the staircase lands somewhere nobody drew. The drawing shows the staircase perfect. The metal shows the sum.
A tolerance is how far a part is allowed to be off its perfect size. A stack-up is what happens when the parts are put together: every part is a little off, and the little offs add up. The assembly that fits on screen fails in metal because nobody added them up, and the fix is rarely a tighter tolerance. It is usually a shorter staircase.
What a tolerance actually promises
A dimension on a drawing is a promise with two numbers: the nominal (the perfect size) and the tolerance (how far off the perfect size the part is allowed to be). A part drawn as 40.00 ± 0.20 mm is promised to arrive between 39.80 and 40.20 mm. The promise is not that the part will be 40.00. The promise is that it will be inside the band, and anywhere inside the band is a good part.
That is the first thing most drawings get wrong: the nominal is not the part. The band is the part. And when five parts go together, five bands go together.
GD&T, Geometric Dimensioning and Tolerancing, is the language for writing that promise properly, standardised as ASME Y14.5-2018. It does two things a plain ± number cannot: it controls geometry, not just size (how flat a face is, how square a hole is to a face, how truly positioned a hole is), and it names the datum, the reference the measurement starts from. A hole is not "±0.1 mm from the edge." A hole is "±0.1 mm from datum A, the face the part sits on." Without a datum, two people measure the same part two different ways and get two different answers, and both are right. The datum is the promise's starting point.
Process capability is the other half of the promise: whether the machine that makes the part can actually hold the band. Cp is the width of the tolerance band divided by six standard deviations of the process (Cp = (USL − LSL) ÷ 6σ, where USL and LSL are the upper and lower specification limits and σ is the process's standard deviation). Cpk is the same idea, adjusted for whether the process is centred on the nominal or drifting to one side. The guides' rule of thumb: aim for Cpk ≥ 1.33–1.67 on critical features, and Cpk ≥ 1.0 may be acceptable on less critical ones. A tolerance tighter than the process can hold is not a specification. It is a wish, and the shop will either charge for it, or quietly ship parts that don't meet it.
The two ways to add them up
Ten parts at ±0.1 mm, two ways to add them up
| Method | Stack | Assemblies that fit |
|---|---|---|
| Worst case: add them all (10 × 0.1) | ±1.00 mm | 100% |
| Root sum of squares | ±0.32 mm | 99.73% at 3σ (0.27% rework) |
Worst case adds the tolerances directly. The assembly tolerance is the sum of the absolute values of every contributor: TG,WC = Σ |ti|. Ten dimensions at ±0.1 mm stack to ±1.0 mm. Worst case guarantees 100% assembly yield with full interchangeability, every part from any source at any time will assemble and function correctly, because the calculation assumes every part lands at its worst limit in the same direction at the same time. The price is that the individual tolerances must be tight, and tight tolerances are expensive.
RSS, root sum square, adds the tolerances statistically. The assembly tolerance is the square root of the sum of the squares: TG,RSS = √(Σ ti²). It assumes each tolerance represents ±3σ (three standard deviations) of a normal distribution, and that the parts vary independently and are centred on their nominals. Ten dimensions at ±0.1 mm stack to ±√(10 × 0.01) = ±0.316 mm, about a third of the worst-case answer. The RSS result means 99.73% of assemblies will fall inside the band, and the remaining 0.27%, about 2,700 per million, will fall outside it and need rework or rejection. For four equal tolerances RSS is half the worst-case value; for nine equal tolerances it is one third. The probability of all ten parts landing at their limits at once is vanishingly small, under 1 in 10 million, per the published guides.
Monte Carlo is the third method, reserved for the problems the first two cannot handle: non-linear chains, correlated parts (two dimensions cut by the same tool, so they drift together), and three-dimensional geometry. It simulates thousands of random assemblies and counts how many fit. It is the right tool for a complex chain, and overkill for a straight one.
The assumptions are the whole game. RSS is only honest when the parts are statistically independent, when each supplier's process genuinely fills its tolerance band symmetrically at the assumed capability, and when the assembly relation is linear. If a process runs off-centre, or an operator sorts parts so only one side of the tolerance ships, the 0.27% figure no longer applies. And the mean shifts, the systematic drift of a process away from nominal, are stacked arithmetically, not statistically: that is the method of Scholz's tolerance-stacking review and the NISTIR 6223 report, which the guides cite as the anchor for both methods. Most bad stack-ups fail on setup, not on maths: a missed dimension, a wrong direction, or a statistical method used on a process that is not centred.
The cost of a tolerance tighter than the process can hold
Tighter tolerances cost more, fast: relative cost for general metals (vendor CNC guides, labelled as such)
| Tolerance | Class | Relative cost |
|---|---|---|
| ±0.10 mm | Commercial (ISO 2768-m) | 1.0× |
| ±0.05 mm | Standard precision | 1.3 to 1.6× |
| ±0.01 mm | High precision, temperature-controlled | 2.0 to 2.8× |
| ±0.002 to 0.005 mm | Ultra-precision, grinding | 3.5 to 5× |
Manufacturing cost against tolerance follows a nonlinear curve: fairly flat for loose tolerances, rising slowly, then climbing sharply as tolerances tighten toward the process limits. The published CNC cost gradient for general metals (aluminium, stainless, tool steel), a vendor guide, labelled as such, reads: ±0.10 mm is the commercial general tolerance (ISO 2768-m default) at 1.0× baseline, on a standard 3- or 5-axis CNC in a regular shop; ±0.05 mm is standard precision at 1.3–1.6×; ±0.01 mm is high precision at 2.0–2.8×, needing a temperature-controlled workshop, high-feed finishing, frequent probing, and premium carbide tools; ±0.002–0.005 mm is ultra-precision at 3.5–5×, needing a thermally stabilised cleanroom CNC, grinding or honing, and 100% CMM inspection.
Tightening a tolerance from ±0.1 mm down to ±0.01 mm doubles or triples the part price. And the same guide's claim, labelled as a vendor claim, is that proper tolerance zoning, GD&T usage, stack-up calculation, and process-aligned tolerance matching can cut component costs by 25–40% while fully locking the critical fit.
A 2023 audit of 17 Tier-1 suppliers serving EU-based EV charging infrastructure OEMs, a consultancy analysis, labelled as such, puts numbers on the same step: identical geometry at ±0.05 mm cost a median $89.40 a unit against $31.20 at ±0.10 mm, a 2.86× premium. And 68% of that premium came not from material or labour but from process overhead: rework (22%), coordinate measuring machine time (31%), and fixture recalibration (15%).
The cycle time rose 37% on average, and the scrap rate on aluminium 6061 rose from 0.8% to 3.4%, a 4.25× rise. At ±0.025 mm the cycle time rose 112% and the scrap rate reached 12.6%. The root cause is the capability maths: at ±0.1 mm, a standard deviation of 0.025 mm satisfies Cp ≥ 1.33; at ±0.05 mm the same σ must drop to ≤0.0125 mm, pushing Cp requirements into the 1.67–2.0 range, achievable only with thermally stabilised machining centres, air-bearing spindles, and real-time probe feedback.
And a separate comparison guide's rule, labelled as such: over-tightening tolerances beyond functional requirements can increase CNC machining costs by 200–300%.
The process capability ranges, as guidelines to verify with your own supplier: general CNC turning and milling holds ±0.05–0.2 mm; precision grinding holds ±0.005–0.02 mm. The rule the guides all repeat is the same: don't assume default machine capability, ask your supplier for σ or Cpk data on the specific feature, process, and material. The tolerance you can afford is the tolerance your process can hold.
The failure modes: how a stack-up actually fails
The missed dimension. The chain that decides the gap has five links, and the drawing accounts for four. The fifth, the gasket's compressed thickness, the board's solder-mask height, the adhesive's bond line, is the one that eats the margin. The loop diagram is the foundation of the analysis for exactly this reason: trace every dimension that controls the gap, in order, from one end of the chain to the other, and check the direction of each one. A dimension that subtracts when you thought it added is a sign error, and a sign error is a stack-up that passes on paper and fails on the line.
The wrong method. RSS used on a process that is not centred, a supplier whose parts all run to the high side of the band, because their tool wears in one direction, gives a 0.27% figure that no longer applies. Worst case used on a high-volume assembly forces tolerances the process cannot hold, and the shop either charges for them or ships parts that don't meet them. The method must match the production reality: worst case for safety-critical fits, very low volume, and zero-defect interfaces; RSS for high-volume production where a small rejection rate is economically preferable to tighter tolerances, and where the independence assumption is valid, parts from many different batches, not one batch cut by one tool.
The tolerance that is tighter than the process. The drawing says ±0.01 mm on a feature the shop holds at ±0.05 mm. The shop quotes the ±0.01 mm price, two to three times the part cost, or it quotes the ±0.05 mm price and ships parts that fail first-article inspection. Either way the assembly yield is a coin toss, and the coin is the supplier's capability, not the drawing.
The chain that is too long. Five contributors at ±0.2 mm stack to ±1.0 mm worst case and ±0.45 mm RSS. Three contributors at the same tolerances stack to ±0.6 mm worst case and ±0.35 mm RSS. The chain length is what drives the stack, and the chain length is a design decision, not a law of nature. That is the demo.
Example Labs is a made-up company we use for worked examples.
DEMO
This product is made up. Call it the Example Labs field sensor hub again: the fence-post unit from Articles #6 and #7, the weather sensor, the event camera from Article #2, the Kalman filter, the PID loop, now in its 100 × 60 × 40 mm enclosure, with the camera looking out through the enclosure's window. The demo's arithmetic is the article's own, verified before writing; the cost figures are the published guides cited above, labelled as vendor figures.
Step 1, The gap that matters. The critical dimension is the gap between the top of the camera module and the inside of the lid. Nominal gap 2.00 mm. It must stay between 1.50 and 2.50 mm: below 1.50 the lens touches the lid, above 2.50 the camera's focus alignment drifts. The chain has five links, and the nominals add up: enclosure internal depth 40.00 mm, minus the board assembly 6.60 mm, minus the camera module 12.00 mm, minus the gasket's compressed thickness 2.00 mm, minus the lid's inner boss 17.40 mm, which leaves 2.00 mm. The tolerances, each bilateral, are: enclosure ±0.20 mm, board assembly ±0.15 mm, camera module ±0.25 mm, gasket ±0.20 mm, lid boss ±0.15 mm.
Step 2, Worst case: it fails. Add the tolerances directly: 0.20 + 0.15 + 0.25 + 0.20 + 0.15 = ±0.95 mm. The gap runs from 1.05 to 2.95 mm, outside the 1.50–2.50 mm window on both sides. Worst case says the assembly will not fit. The drawing that shows it fitting is the drawing of the nominals, and the nominals are not the parts.
Step 3, RSS: it passes. Square, add, square-root: √(0.20² + 0.15² + 0.25² + 0.20² + 0.15²) = √(0.1875) = ±0.43 mm. The gap runs from 1.57 to 2.43 mm, inside the window, at 3σ, which means 99.73% of assemblies fit and 0.27% fall outside. At 1,000 units a year, that is about 3 units that need a shim or a re-seat. The same parts, the same tolerances, and the answer flips, because the method changed.
Step 4, What each answer costs. The worst-case answer costs a redesign of the tolerances: to get the worst-case sum under the ±0.50 mm the window allows, nearly every tolerance in the chain must tighten, and tightening from ±0.1 mm toward ±0.01 mm doubles or triples the part price, with the scrap rate climbing from 0.8% toward 3.4% and beyond. The RSS answer costs about 3 units a year of rework, a shim, a re-seat, ten minutes each. The honest question is which one the product can afford: the tighter tolerances, or the three shims.
Step 5, The redesign that beats both. Shorten the chain. Mount the camera directly to the lid, and the gasket and the lid boss leave the chain, the camera's position is set by the lid it is mounted to, not by the stack beneath it. The chain drops from five contributors to three: enclosure ±0.20, board assembly ±0.15, camera ±0.25. Worst case becomes ±0.60 mm (gap 1.40–2.60 mm, still outside the window, but closer), and RSS becomes ±0.35 mm (gap 1.65–2.35 mm, inside the window with more margin than before, and no tolerance tightened anywhere). The chain length is what drives the stack. Shorten the chain and the problem shrinks faster than any tightening can.
The worked example's camera gap: nominal 2.00 mm, must stay between 1.50 and 2.50 mm. Same five parts (±0.20, 0.15, 0.25, 0.20, 0.15 mm), two methods.
| Method | Stack | Gap range | Result |
|---|---|---|---|
| Worst case (add them) | ±0.95 mm | 1.05 to 2.95 mm | Fails |
| Root sum of squares (3σ) | ±0.43 mm | 1.57 to 2.43 mm | Passes; about 3 units in 1,000 need a shim |
The decision test
Four questions. They sort the stack-ups that are honest from the products that need a tolerance review first:
- Has the stack-up been run, the gap named, the chain traced, every contributor listed with its direction? No, the assembly yield is a coin toss, and the line will tell you first. Yes, the fit is a calculation, not a hope.
- Does the method match the production reality, worst case for safety-critical and low-volume, RSS for high-volume with independent, centred processes? No, the method is a wish, and the 0.27% figure is a fiction. Yes, the yield prediction is honest, and the rework is budgeted.
- Can the process hold the tolerances on the drawing, Cp and Cpk from the supplier, on the specific feature, not a default? No, the tolerance is a wish, and the shop will charge for it or ship parts that don't meet it. Yes, the tolerance is a specification, and the price is real.
- Is the chain as short as it can be, every contributor earning its place, and the longest chains redesigned rather than tightened? No, the stack is paying for links that don't need to be there, and the tightening is paying for the chain. Yes, the design is doing the work, not the tolerance.
If the answers are yes, yes, yes, yes, the stack-up is honest, and the review is a final check, not a rescue. If any answer is no, the job is a tolerance review, and the honest thing is to do it before the tooling is cut, not after the first hundred units come back from the line. And if the answer to question 4 is "we tightened everything instead", the problem is the chain, not the tolerances, and the review starts by asking which links can leave.
"But we tightened the tolerances", and two more fair doubts
"We tightened the tolerances." Tightening is the most expensive way to fix a stack-up, and it is the first thing a team reaches for. The cost curve says why: tightening from ±0.1 mm to ±0.01 mm doubles or triples the part price, and the scrap rate climbs with it. The redesign, a shorter chain, a datum moved to the face that matters, a part that locates on the feature it actually touches, costs nothing at design time and shrinks the stack faster than any tightening can. Tighten last, not first.
"RSS is just gambling." RSS is gambling only if the assumptions are not checked. If the parts are independent, centred, and normally distributed, RSS is the honest prediction, 99.73% inside the band, 0.27% outside, and the 0.27% budgeted as rework. The gamble is using RSS on a process that is not centred, or on parts from a single batch cut by a single tool, and calling the 0.27% a fact. The review's job is to check the assumptions, not to bless the method.
"Our supplier holds ±0.01 mm." Then ask for the capability study: the σ and the Cpk, on the specific feature, on the specific material, on the specific machine, not a default from a catalogue. A supplier that holds ±0.01 mm on one feature may hold ±0.05 mm on the next, and the drawing that assumes the first number for the second feature is a wish. The review reads the capability study, not the sales sheet.
Sources
- Sekkei Tech, "Tolerance Stack-Up Analysis: Worst Case vs Statistical Methods" (June 2026)
What we used
Loop diagram method; worst-case TG,WC = Σ|ti|, guarantees 100% yield with full interchangeability; RSS TG,RSS = √(Σ ti²) with each tolerance as ±3σ; 10 dimensions at ±0.1 mm: WC ±1.0 mm, RSS ±0.316 mm; RSS means 99.73% inside, 0.27% outside; RSS appropriate for high-volume production where a small rejection rate is economically preferable to tighter tolerances, and where the independence assumption is valid
- Flying Chip Factory, "Tolerance Stack-Up: How to Calculate It and Avoid Mistakes" (November 2026)
What we used
The stack-up answers one question: with every part at its tolerance limits, does the critical gap or fit still work; most bad stack-ups fail on setup: a missed dimension, a wrong direction, or a statistical method used on a process that is not centred; RSS with inflation factors for nonnormal distributions and stacking with mean shifts (Scholz / NISTIR 6223); if a process runs off-centre or parts are sorted, the 0.27% figure no longer applies
- HLH Sheet Metal, "Tolerance Stack Up for Engineers: Equations and a 4 Part Example"
What we used
Worst case for safety-critical or zero-defect interfaces, RSS when contributing processes are independent and centred, Monte Carlo for nonlinear, correlated, or 3D problems; ASME Y14.5 and NIST guidance anchor the methods; Scholz's review: mean shifts stacked arithmetically, residual variability statistically; ASME Y14.5-2018 (Dimensioning and Tolerancing)
- Nova Calculator, "Tolerance Stackup Calculator" (August 2025)
What we used
RSS = √(t1² + … + tn²), assumes normal distribution and independent tolerances; for statistical analysis multiply RSS by Z/3; for four equal tolerances the RSS result is half the worst-case value, for nine equal tolerances one third; 3σ = 99.73% inside (about 2,700 nonconforming per million), 2σ leaves roughly 4.6% outside, 3.89σ about 99.99%; RSS valid only when contributors are independent, each process genuinely fills its band symmetrically, and the assembly relation is linear
- BlackRock Engineering, "Tolerance Allocation: Cost, Quality & Manufacturing"
What we used
Cost vs tolerance is nonlinear: flat for loose tolerances, then rising slowly, then climbing sharply toward process limits; Cp/Cpk definitions; critical features Cpk ≥ 1.33–1.67, less critical Cpk ≥ 1.0; CNC turning/milling general ±0.05–0.2 mm, precision grinding ±0.005–0.02 mm; don't assume default machine capability, ask the supplier for σ or Cpk data on the specific feature/process/material
- Zorapid, "Tolerance Design Guide to Cut Industrial Component Costs" (vendor guide, labelled as such)
What we used
Cost rises exponentially as tolerance shrinks; CNC cost gradient for general metals: ±0.1 mm 1.0× baseline (ISO 2768-m), ±0.05 mm 1.3–1.6×, ±0.01 mm 2.0–2.8×, ±0.002–0.005 mm 3.5–5×; tightening ±0.1 mm to ±0.01 mm doubles or triples part price; proper zoning, GD&T, stack-up, and process-aligned matching can cut component costs 25–40% (vendor claim, labelled as such); odds of all features hitting extreme limits simultaneously under RSS are under 1 in 10 million
- TradeNexus Pro, "CNC Machining Tolerances: Why ±0.05 mm Costs 3× More"
What we used
(April 2026, consultancy analysis, labelled as such), 2023 audit of 17 Tier-1 suppliers: median unit cost $89.40 at ±0.05 mm vs $31.20 at ±0.10 mm (2.86× premium); 68% of the delta from process overhead (rework 22%, CMM time 31%, fixture recalibration 15%); cycle time +37% average, scrap 0.8% → 3.4% (4.25× rise); at ±0.025 mm cycle time +112%, scrap 12.6%; at ±0.1 mm a σ of 0.025 mm satisfies Cp ≥ 1.33, at ±0.05 mm σ must drop to ≤0.0125 mm, pushing Cp to 1.67–2.0
- CISRI-ATM, "MIM vs CNC Machining vs Precision Casting: A Complete 2026 Comparison Guide"
What we used
Over-tightening tolerances beyond functional requirements can increase CNC machining costs by 200–300%; apply GD&T only where functionally necessary