Fiducials are not decoration. They are the only optical reference the pick-and-place machine trusts when it maps the FR4 panel to its camera coordinate system. If FR4 PCB fiducial design is sloppy, the machine still runs. It just places every fine-pitch package from a slightly wrong origin.
Most placement errors blamed on "the assembler" start as a missing local mark, a pad covered by silkscreen, or three global marks sitting almost in a straight line. PCB fiducial placement is a layout rule, not a CAM afterthought.
What a Fiducial Actually Does on the Line
The camera looks for a high-contrast circle: bare copper (or finish metal) against soldermask. It finds the centroid, then computes offset, rotation, and stretch of the board or the local region. Global marks correct the whole panel. Local marks correct the neighborhood around a tight-pitch part.
FR4 itself shrinks and warps a little through etching, baking, and reflow. Panel tooling holes tell the conveyor where the board is. They do not tell the vision system where the copper pattern landed. That is the fiducial's job.
SPI and AOI use the same marks. If the mark is usable for placement, it is usually usable for inspection. If it is dirty, clipped by routing, or sitting under a shadow from a tall part, both processes lose the reference.
Global Versus Local Fiducials
Global fiducials belong to the board or the panel. Three is the useful number. Two marks give translation and rotation. They cannot separate stretch from rotation cleanly. Three marks, not colinear, give the machine an affine fit. That matters on larger FR4 panels where etch and bake stretch is not uniform.
Place them far apart. One in each of three corners is the usual pattern. Do not put all three along one edge. Do not put them in a perfect isosceles layout so tight that a small camera error still looks "good." Distance is what makes the transform stable.
Local fiducials belong to a package. Use them next to BGAs, fine-pitch QFNs, 0.4 mm and 0.5 mm QFPs, and any connector whose pins are tighter than the global correction can guarantee. Two locals per device is common—diagonally opposite. One local is better than none, but it only corrects translation in that pocket.
Panel fiducials and board fiducials are not interchangeable. If the assembler runs boards in a pallet after depaneling, the board-level marks must survive. If they only exist on the rails that get routed off, the next process has nothing to see.
Size, Clearance, and Finish That the Camera Can Lock Onto
A 1.0 mm copper diameter is the default that almost every SMT line accepts. 0.5 mm works on some cameras and fails on others. 1.5 mm is easy to see and wastes board area. Stay at 1.0 mm unless the assembler specifies otherwise.
The soldermask opening must be larger than the copper. A 2.0 mm mask opening on a 1.0 mm pad (0.5 mm clearance all around) keeps mask from creeping onto the metal and breaking the circle. No silkscreen in that keep-out. No via, no trace, no plane edge, no text.
Keep other copper and graphics at least 2 mm away from the pad edge, 3 mm if the board is crowded and contrast is already poor. The vision algorithm expects a clean annulus of mask around a round metal disk. A trace kissing the pad turns the blob into a comma. The centroid shifts.
Surface finish matters more than designers admit. ENIG gives a flat, bright circle. Immersion silver is similar. HASL can leave a dome or an uneven edge; some lines still accept it, some hunt for the centroid and miss. OSP is darker and depends on the mask color. If the board is HASL and the pitch is tight, do not argue about finish after the first placement reject. Change the mark contrast or the finish on that job.
Same layer as the parts being placed. A bottom-side only fiducial does nothing for a top-side camera. Double-sided SMT needs marks on both sides, or a process that flips and re-references.
Where PCB Fiducial Placement Goes Wrong Near the Outline
Keep global marks off the routing channel and away from clamps. A 3–5 mm setback from the finished outline is a practical minimum on most FR4 jobs. Closer than that, the router nicks the pad or the conveyor clamp covers it. Breakaway tabs are worse: the mark looks fine in Gerber and disappears when the tab is snapped.
Do not hide a mark under a tall connector, a shield can, or a heat sink that is already on the board from a first-side pass. Second-side vision cannot see through hardware.
Do not put the only usable marks inside a region that will be covered by a conformal coat inspection fixture or a press-fit block. If a later process still needs optical alignment, leave a pair that survives.
How SMT Machines Read the Mark—and What They Reject
The algorithm wants a round, high-contrast target of expected diameter. It measures area, circularity, and brightness. An etched pad that came out oval fails circularity. A pad with soldermask on one edge fails brightness on that side and pulls the centroid toward the clean metal.
Lighting is top-down. Glossy silk next to the pad creates glare. A via in the keep-out creates a dark hole the camera may treat as part of the shape. Copper pour too close reduces the dark ring the software uses as an edge.
Global correction is applied first. Local correction is applied on top for that package. If the locals are 8 mm from a 0.4 mm-pitch QFP, they still help. If they are 40 mm away, they are just extra globals and do not fix local warp between the mark and the pads.
Put locals in the same copper layer as the device pads, close to the package, outside the paste and courtyard. Inside the paste stencil aperture is a different kind of mistake: the mark gets printed or smeared and the camera sees solder, not copper.
Design Errors That Show Up After Gerber Release
Only two globals, colinear. The machine accepts the board and then fights stretch on a long panel.
Fiducial drawn as a flashed pad with soldermask expansion set to zero. Mask covers the edge. Contrast dies.
Silk reference designators parked on top of the mark because the courtyard was empty in the library.
Local marks omitted on the one BGA that needed them, because "we already have globals." Globals do not correct local coupon stretch next to a large package.
Marks placed only on the panel rails. Board-level assembly after scoring has no reference.
Different diameters mixed on the same side with no note to the assembler. The vision job is written for one size. The second size is ignored or misread.
Fiducial sitting in a copper pour with only a tiny mask dam. The pour becomes the background and the circle disappears.
A Placement Rule Set That Survives SMT
Three global fiducials, 1.0 mm copper, 2.0 mm mask opening, 3 mm keep-out, not colinear, set back from the outline and from clamps. Local pair next to every fine-pitch or large BGA, on the same side as the part, outside paste. Same diameter everywhere on that side. No silk, no vias, no plane edge in the keep-out.
FR4 PCB fiducial design is finished when the assembler can write one vision job and run both sides without hunting for a usable circle. Everything after that is yield. If the mark is wrong, the machine will still place parts. It will just place them to a reference that does not match the copper.
