Industries / Technology / Software and SaaS / Engineering design software
Drawing and 3D modeling
Drawing and 3D modeling is the work of describing a part's exact shape on a computer, first as lines on a screen and now as solid models that the software can measure, change and turn into drawings.
6
Milestones achieved, from 1964 to 2001.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1964
A carmaker designs on screen
General Motors Research Laboratories put a laboratory for "graphical man-computer communication" into operation in 1963 and described it in papers at the 1964 Fall Joint Computer Conference in Detroit. The system, called DAC-1 for "Design Augmented by Computers", joined a large IBM computer, a display console and a light pen, so designers could draw, rotate and change car shapes on screen instead of building physical models. The Computer History Museum says IBM completed the DAC-1 software for GM in 1964 and that GM used it until the late 1960s, when a new president mothballed the system and broke up its team. Sources disagree on the computer: the museum names an IBM 7090, GM's own papers an IBM 7094.
Lessonan early design system could be shelved by one change of leadership, however much it had cost to build.
Credit
General Motors Research Laboratories, including Edwin L. Jacks; IBM
1985
A leader misses workstations
Computervision, one of the largest CAD vendors of the 1970s, sold its drafting software on minicomputers it built itself. In late 1982 it told computer makers it would move to standard networked workstations, chose Sun Microsystems in 1983, and then spent until 1986 porting the five to six million lines of its CADDS 4X software, by the account of the CAD historian David Weisberg. In 1985 its revenue fell to $441 million with an $81 million loss, and it laid off nearly 2,000 people by the end of that year. Prime Computer launched a hostile takeover bid for it in December 1987.
Lessona design-software company that also built its own computers fell behind once cheaper standard hardware could run the same work.
Credit
Computervision
1988
Models rebuilt from history
Parametric Technology Corporation (PTC) began commercial shipments of Pro/ENGINEER in January 1988, after Deere & Company bought the first two copies in September 1987. It built parts from features, such as holes and cuts, sized by dimensions, and it recorded every step in a "history tree": change a dimension and the software replayed the steps to rebuild the part. Weisberg writes that each idea had appeared in earlier products, but this was the first time they were grouped in one package, and that PTC shipped nearly 900 copies to about 150 customers within a year.
Lessoneach feature depends on the ones made before it, so large models rebuilt slowly and users had to plan a model's steps to avoid starting over after a big change.
Credit
Samuel P. Geisberg, Parametric Technology Corporation
1989
One kernel, many programs
A geometry kernel is the core code that computes solid shapes, such as joining two solids or rounding an edge, underneath a CAD program's screens. Shape Data, a company in Cambridge, England, built Parasolid as component software that other software firms could use in their own products. McDonnell Douglas bought Shape Data in 1988 and shipped Parasolid inside Unigraphics 7.0 in December 1989. SolidWorks switched to it from Spatial's ACIS kernel before its 1995 release, and Bentley and Intergraph's Solid Edge followed by 1997 (Weisberg). In 2025 Zoo, a newer CAD company, wrote that SolidWorks, Onshape and Shapr3D "still wrap Parasolid".
Lessonat first other vendors were wary of building on a critical piece of software controlled by a competitor.
Credit
Shape Data; McDonnell Douglas and its Unigraphics unit, now part of Siemens
1995
Solid modeling on Windows
SolidWorks unveiled SolidWorks 95 at the AUTOFACT conference in 1995. It was a solid modeler driven by features and dimensions, like Pro/ENGINEER, but it ran only on Windows PCs and looked and worked like Microsoft's own programs, and its drawings stayed linked to the model in both directions. By June 1997 it had sold about 6,000 copies, and Dassault Systèmes agreed to buy it for $300 million, about 12 times its annual revenue (Weisberg). Dassault's own history says it bought SolidWorks to bring "cost-effective, easy-to-use" solid modeling to the mainstream market.
Credit
Jon Hirschtick and the SolidWorks founding team; Dassault Systèmes
2001
A CAD standard merged away
In December 1995 Ford signed a $200 million contract that made SDRC's I-DEAS its main design software, and by April 1998 it had installed 3,000 seats of I-DEAS and 2,600 of SDRC's Metaphase data management software (Weisberg). On May 23, 2001 EDS agreed to buy SDRC for about $950 million and merged it with its Unigraphics business. The plan was to bring I-DEAS and Unigraphics closer with each release until they were "one and the same"; the combined product, NX, was announced in October 2002, and I-DEAS models were to be rebuilt in it from their history and feature data rather than translated.
Lessona company's choice of design software can be undone by its vendor's sale, and moving years of models to another system meant rebuilding them.
Credit
SDRC; EDS and its UGS unit
Watch for
A GPU geometry engine
Zoo says it wrote its own geometry engine for Zoo Design Studio. In its telling the field was rebuilt with Parasolid in 1988 and ACIS in 1989, and the last major commercial engine to reach mainstream use came "well over a decade ago". It says it recast surface-to-surface intersection, one of a kernel's hardest jobs, as a problem that runs in parallel on graphics processors (GPUs), and that it will "publish detailed benchmarks as further optimizations land".
Achieved whenan outside party runs the same public set of parts through Zoo's engine and through Parasolid or Open CASCADE and publishes the success rates, accuracy and times.
Claimed by
Zoo, May 2025
Open
01
Topological naming stress test
A public test suite and script that take parametric part models, change each driving dimension up and down in steps, rebuild each model, and sort every rebuild into three bins: correct, failed with an error, or rebuilt but wrong (a later feature attached to the wrong face). The underlying fault is called topological or persistent naming: after an edit changes a model's faces and edges, the software must find "the same" face again so the features built on it stay put.
The briefCredit · unclaimed02
One test for every kernel
An open test suite that gives the same inputs to every geometry kernel whose licence allows published results, starting with three open ones, Open CASCADE, NVIDIA's SMLib and Truck, and publishes one table. For each file a kernel can import and each scripted operation (join, cut, round an edge, hollow out, offset) it records whether the kernel succeeded, whether the result is a valid closed solid, and how far its volume and surface area differ from the others'.
The briefCredit · unclaimed
Open 01 · the brief
Topological naming stress test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test suite and script that take parametric part models, change each driving dimension up and down in steps, rebuild each model, and sort every rebuild into three bins: correct, failed with an error, or rebuilt but wrong (a later feature attached to the wrong face). The underlying fault is called topological or persistent naming: after an edit changes a model's faces and edges, the software must find "the same" face again so the features built on it stay put.
Why it is still open
David Marcheix wrote in 2006 that "most systems fail during the re-evaluation phase when various kind of topological changes occur". FreeCAD 1.0, released on 19 November 2024, listed "the inclusion of the topological naming problem mitigation code" as its most notable change. A 2016 survey of CAD model quality tools found that no quantitative comparison against a common benchmark had been made, and no public rebuild-failure benchmark turned up in a search for this file.
Built with what exists
FreeCAD 0.21 and 1.x, which run without a screen from Python scripts, for a before-and-after comparison; a commercial system only where its licence allows published results (Onshape's terms, for example, forbid analysing its service "for purposes competitive to Onshape"); parametric part files from public open-hardware repositories; Open CASCADE to measure each rebuilt solid. A coding agent such as Inkling 2 Small (see the Note "Checked without being asked") can write and debug the harness, and each case can be run many times to separate real faults from flaky ones (see the Note "All twenty times").
First step
take 20 FreeCAD part files from public repositories, write a script that changes each sketch dimension by plus and minus 20 percent, and log which rebuilds fail in FreeCAD 0.21 and 1.x.
Done looks like
at least 100 public models and 2,000 dimension changes run in FreeCAD 0.21, FreeCAD 1.x and, where its licence allows, one commercial system; the share of rebuilds in each bin published for each program; at least 200 "correct" rebuilds checked by a person from rendered images; and the suite released so vendors can rerun it.
Open 02 · the brief
One test for every kernel
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open test suite that gives the same inputs to every geometry kernel whose licence allows published results, starting with three open ones, Open CASCADE, NVIDIA's SMLib and Truck, and publishes one table. For each file a kernel can import and each scripted operation (join, cut, round an edge, hollow out, offset) it records whether the kernel succeeded, whether the result is a valid closed solid, and how far its volume and surface area differ from the others'.
Why it is still open
The industry's public STEP AP242 benchmarks test commercial CAD programs' file translators and publish "only those results … which are confirmed by the respective IT vendors"; NIST's 2015 tests covered four major commercial CAD systems. Two new kernels appeared after those tests, Zoo's in May 2025 and NVIDIA's open-sourced SMLib in October 2026, and Zoo says its own benchmarks are still to come.
Built with what exists
Open CASCADE (open source); NVIDIA's USD BRep library, which packages SMLib under the Apache 2.0 licence with C++ and Python interfaces; Truck, a kernel under the Apache 2.0 licence that re-implements exact NURBS solids in the Rust language; the ABC dataset, one million CAD models collected from Onshape, which always includes STEP and Parasolid files and whose models remain their creators' copyright under Onshape's terms; NIST's test-part STEP files, which may be used without restriction; a coding agent to write the scripts for each kernel's interface. Commercial kernels such as Parasolid and Zoo's engine can join only where their makers agree: Onshape's terms of use, for one, forbid analysing its service "for purposes competitive to Onshape".
First step
import 500 ABC STEP files with Open CASCADE, convert the same solids into USD BRep with its Open CASCADE converter, compute each solid's volume and area in both kernels, and list every file where they differ by more than 0.01 percent.
Done looks like
10,000 ABC files and 1,000 scripted operations run through at least three kernels; a public table of success rates, valid-solid rates and cases where volumes differ by more than 0.01 percent; the scripts released and the table rerun on each new kernel release.