CH NEO-ZÜRICH EDITION
WEATHER · OVERCAST 12°C
BLEND OF THE DAY · 07/ROGUE
EST. 2027
THE AEC CYBER MORNING NEWS

PAZ Kaffi

DESIGN · DEMOLITION · CAFFEINE · DISPATCH
EDITION 1117 · 17 November 2026
BROADCAST 04:42 CET
2,400 BROADSHEETS PRINTED
READ TIME · 47 MIN
Floating Point from Zero: Why LV95 Coordinates Step in 25 cm on Screen
SYSTEMS
FRAME · 07:00
17-11-2026

Floating Point from Zero: Why LV95 Coordinates Step in 25 cm on Screen

Why a Rhino model at Swiss LV95 coordinates shimmers: IEEE 754 single precision steps in 25 cm at 2,600,000 m. Model near zero, record the offset.

Read swisstopo’s description of the national grid and the first surprise is that the zero point is not at zero. The old observatory in Bern, the defined origin of Swiss coordinates, sits at E = 2,600,000 and N = 1,200,000 in LV95. That puts every georeferenced Swiss project millions of metres from the number a computer treats as home: a school in Winterthur, a bridge in Graubünden, a housing block in Lausanne. Most days nobody notices. Then a façade starts to shimmer on screen while the model underneath is perfectly sound. The reason is a decision made in the late 1970s about how a machine should hold a number.

This season David Goldberg’s paper What Every Computer Scientist Should Know About Floating-Point Arithmetic (ACM Computing Surveys, March 1991) turns 35. It opens with a sentence that still holds in every architecture office: “Floating-point arithmetic is considered an esoteric subject by many people.” It does not have to be. The whole idea fits on one line. Once you own that line, the 25 cm mystery solves itself.

What it is

What it is: A floating-point number is scientific notation written in binary. It has a fixed count of significant digits, the significand, and an exponent that slides the point left or right. In decimal you already know the move. 6.02 × 1023 and 1.60 × 10−19 spend the same three digits, and only the exponent tells you whether you are counting atoms or charges.

The consequence is the one sentence worth owning: precision is relative, not absolute. You always get the same number of digits, wherever the point lands. Near zero those digits can describe micrometres. At two and a half million, most of them are busy describing the millions, and what remains for the centimetres is coarse.

There is a second, quieter consequence. As Goldberg shows, the decimal 0.1 “in binary has an infinite repeating representation”, the way 1/3 never terminates in decimal. So 0.1 “is exactly representable by neither” of its neighbouring floats. The machine stores the nearest one and moves on. Every slider value you type into Grasshopper passes through that same act of rounding.

Why it works

Why it works: Think of the representable numbers as ticks on a ruler. Between each pair of neighbouring powers of two, the ticks are evenly spaced, and there is always the same number of them. Cross into the next power of two and the spacing doubles. It is a ruler whose graduation stretches with distance from the origin. That gap between neighbours is called the unit in the last place (ulp), and it is the real resolution of your coordinate.

IEEE 754 fixes the digit budget. Single precision (binary32) carries a 24-bit significand, about 7.22 decimal digits. Double precision (binary64) carries 53 bits, about 15.95 digits. Now do the geometry. An easting of 2,600,000 m lies between 221 and 222. The float32 gap there is 221 / 223 = 0.25 m. A PAZ NumPy test run with np.spacing walks the whole ruler:

  • at 10 m: 0.00000095 m
  • at 1,000 m: 0.06 mm
  • at 50,000 m: 3.9 mm
  • at 1,200,000 m (LV95 north): 0.125 m
  • at 2,600,000 m (LV95 east): 0.25 m

Type float32(2,600,000.123) and it comes back as 2,600,000.0. The 123 mm have nowhere to live. In float64 the gap at the same easting is 0.00000000047 m, well under a nanometre. Switching units will not rescue you, because the ratio is unit-independent. In millimetres, 2.6 · 109 mm steps by 256 mm. Only moving the origin changes the answer.

Single precision halves the memory per vertex and keeps graphics hardware fast. The price is a resolution that grows with distance from zero instead of staying a fixed length. Neither format is wrong. Each is a trade between bits and reach, and the designer chooses where the origin sits.

←TODAY: swisstopo defines Bern’s zero point at E 2,600,000 / N 1,200,000, and a float32 vertex there steps in 25 cm.
→3012: Zurich-3012 archives can still rebuild today’s sites because someone wrote the offset down instead of trusting the file.
Fulcrum: Precision is measured from the origin, so the origin is a design decision.

Origins

Origins: Before 1985, each computer line had its own arithmetic. The same program could return different answers on different machines, and a careful numerical analyst had to learn every dialect. The person who did most to end that was William Kahan, born in Toronto in 1933 and later professor at UC Berkeley. Wikipedia’s page on IEEE 754 calls him “the key designer” of the standard. Kahan himself describes the route through silicon: “In 1977 those features were designed into the Intel 8087 to serve the widest possible market.” Talking to Jack Woehr in Dr. Dobb’s Journal in November 1997, he said the 8087 work “became the basis for the IEEE standard, at least all the rational and algebraic operations.”

IEEE 754-1985 covered binary arithmetic only. A seven-year revision produced the 2008 edition, which added decimal formats, and a minor revision followed in July 2019. In 1989 Kahan received the Turing Award “for his fundamental contributions to numerical analysis.” His own summary of why a standard matters reads like a definition of an academy: “Civilizations are designed to give you the benefits of others’ experience without having to relive it.”

Kahan is the name we know. Around him stood committee members, chip engineers and library writers whose names never appear on a render. Yet every wall length in Archicad and every control point in Rhino passes through rounding rules they argued over line by line. That labour is invisible precisely because it works. It deserves to be said out loud.

One aside shows the same mechanism outside geometry. A Patriot battery at Dhahran had been running for 100 hours on 25 February 1991. A time counter that kept growing lost its fine digits the way a coordinate does, and the drift reached one-third of a second. GAO report IMTEC-92-26 names it a “Software Problem”, and 28 soldiers were killed. The lesson concerns counting: any quantity that keeps growing away from its zero spends its digits on magnitude.

In practice

In practice: Rhino keeps its geometry in double precision, so the model data at LV95 coordinates is sound. The McNeel wiki is precise about where the trouble lives. Objects “extremely far (millions of units) from the world origin suffer from display and precision problems” because of “the polygon mesh vertex locations being represented by a single precision number.” The shimmering edge is the display mesh landing on a 25 cm ruler, not the NURBS surface. McNeel’s own documented remedy is pure geometry. Build the work near the origin, then use Insert to place it “as a Linked and embedded block into the original at the distant location.” The engineering stays local, and the georeference becomes one rigid translation.

PAZ has covered the neighbouring problem before, from the decision side. In our piece on Shewchuk’s robust orientation test, a “which side?” question goes wrong when rounding meets near-collinear points. That piece shows the logic of a test failing. This one shows the screen. Both come from the same ruler with stretching ticks.

Atelier: A Swiss studio receives the survey in LV95, the point cloud in LV95 and the cadastral plan in LV95, and each team quietly imports at full coordinates. Offices that adopt AI-assisted modelling feel this sooner, because a generated geometry inherits whatever origin it was handed. Monday move: pick one round local origin near the site, such as the nearest whole-kilometre LV95 point, and write its E/N offset into the Rhino file’s Notes before anyone models a single wall.

From where I write, in the late 2070s, the files that survived were rarely the elegant ones. They were the ones where someone had written down the offset, the units and the precision, so a stranger could rebuild the site from the principle instead of guessing at the file.

Hack

Hack: Measure the gap between neighbouring numbers at your own site coordinates, in both precisions, before the screen measures it for you. Paste these four lines into any Python with NumPy, including Rhino 8’s script editor if NumPy is installed. The loop prints the float32 and float64 spacing from 10 m out to LV95’s easting. The last line shows a 123 mm offset vanishing. Replace the list with your project’s actual E and N values, then run it again with the local offset subtracted. That second run is the number your display mesh will actually live with.

import numpy as np
for d in [10, 1_000, 50_000, 1_200_000, 2_600_000]:
    print(d, "m  float32 gap:", np.spacing(np.float32(d)), "  float64 gap:", np.spacing(np.float64(d)))
print(np.float32(2_600_000.123))   # 2600000.0 — the 123 mm are gone

The digits are fixed, so distance from the origin eats precision. Open your current LV95 project, read its coordinates, and decide where its zero should be today.

Sources & Further Reading

FILED FROM
CO-SIGNERS
PAZ Academy
CONFIDENCE
HIGH
REPRINTS
© PAZ - PARAMETRIC ACADEMY ZURICH · ALL RIGHTS RESERVED

PAZ Kaffi · multidisciplinary editorial, led by PAZ Academy

⚑ REPORT AN ERROR · SUBMIT A CORRECTION
◂ BACK TO FRONT PAGE · PAZ KAFFI

© 2026 PAZ Academy.