What is a Geometric Kernel?

What is a Geometric Kernel?

A geometric kernel (you will also hear "geometry kernel", "modeling kernel" or just "the kernel") is the software library at the bottom of every serious CAD system that knows what a solid actually is. It creates, edits and validates precise 3D geometry: B-rep solids, exact NURBS surfaces, Boolean operations, fillets, offsets, mass properties. Everything you see in CATIA, NX, Creo, SolidWorks or Fusion 360 is a user interface politely asking the kernel to do the math.

That is the textbook answer. The more useful one is this: the kernel is the most strategic component of the CAD industry that almost no end user has heard of. There are only a handful of them in the world. They are brutally expensive to build, nearly impossible to replace once a product is built on one, and several competing CAD vendors license the same kernel from each other. Every interoperability headache you have ever had with a STEP file traces back to this layer.

I wrote a fifteen-chapter series, Kernel Wars, on exactly this topic. This page is the short version, with pointers into the long one.

Where the kernel sits: anatomy of an MCAD application

Think of a mechanical CAD application as four concentric layers. The diagram above, from Chapter 1 of Kernel Wars, is the mental model I use for every vendor conversation.

1. The geometric modeling kernel (core engine). Solid and surface modeling in boundary representation, with exact surface types including NURBS. Booleans, blends, shelling, offsets, and the topological bookkeeping that keeps a solid watertight after each edit. This is the layer the rest of this page is about.

2. Part and assembly modeling. Two tightly coupled sub-layers sitting directly on the kernel. Part modeling adds the parametric feature history (extrude, hole, fillet), sketches with geometric and dimensional constraints, direct modeling where parametrics get in the way, and the constraint solvers that make it all hold together. Those solvers are often licensed separately from the kernel: Siemens' D-Cubed DCM sits inside a lot of CAD tools that are not Siemens products. Assembly modeling handles mates, kinematics, product structure, interference checking and the lightweight representations that keep a 100,000-part assembly usable. This is where design intent lives, and it is where most of what you think of as "the CAD system" actually happens.

3. Visualization and tessellation. The kernel's exact geometry gets converted into triangle meshes for the viewport and handed to OpenGL, Vulkan or a toolkit like HOOPS. What you see on screen is an approximation. The kernel holds the truth.

4. The application layer. The UI, the CAM module, the simulation add-ons, the generative design workflow. The part you actually pay for. Generative design is worth a pause here: it is essentially a program that hammers the kernel API thousands of times per study. When people say a CAD tool "has AI", the interesting question is how well its kernel copes with being used as a service.

Meshing for finite element analysis straddles layers 3 and 4. It needs the exact geometry from the kernel, not the display mesh, which is why simulation quality is quietly tied to kernel integration quality.

The kernels you will actually meet

Kernel	Owner	Runs inside
Parasolid	Siemens Digital Industries Software	Siemens NX, Solid Edge, SolidWorks, Onshape and 200+ other licensees
CGM	Dassault Systèmes	CATIA V5 and 3DEXPERIENCE; also licensed to third parties through Spatial
Granite	PTC	Creo (the Pro/ENGINEER engine, given a public name in 2014)
ShapeManager	Autodesk	Inventor, Fusion 360 (a fork of ACIS taken in-house in 2001)
ACIS	Spatial, a Dassault Systèmes company	AutoCAD, BricsCAD and a long tail of mid-tier and specialist tools
Open CASCADE	Open Cascade SAS	FreeCAD, plus research, simulation pre-processing and niche commercial tools
C3D	C3D Labs, spun out of ASCON in 2012 and led by Oleg Zykov	KOMPAS-3D and Renga (ASCON), plus a growing set of third-party licensees

Two things jump out of that table. Parasolid is everywhere, including inside products owned by Siemens' direct competitors. And Dassault Systèmes owns two kernels while its own midmarket flagship runs on a third one belonging to its rival. That is not an accident. It is the whole story of the kernel wars.

The full lineage, with acquisitions and forks drawn out, is on the interactive Kernel Map.

Proprietary or licensed: the fork in the road

Every kernel owner faced the same choice: keep the engine for your own products, or sell it to everyone including your competitors.

PTC kept its engine to itself. The parametric geometry core behind Pro/ENGINEER was the first fully parametric engine on the market and an instant success. PTC only put a public name on it, Granite, in 2014, and the attempt to sell APIs around it generated little interest. Dassault took a hybrid path: CGM stayed the private heart of CATIA, while the Spatial acquisition gave it ACIS to sell to the rest of the market.

Unigraphics, later UGS and then Siemens, went the other way and commercialized Parasolid from the start. Tony Affuso, who ran that business for three decades, told me the belief was simple: the kernel is a commodity and the competitive differentiation is the application on top. So they ran a separate "level playing field" toolkit group that gave every licensee, competitors included, the same kernel upgrades on the same terms. That group is Siemens PLM Components today, with well over 200 licensees.

The consequence is the strangest fact in CAD: SolidWorks, the product that made Dassault Systèmes a midmarket power, runs on Siemens' kernel, and has since day one. Chapter 3 tells that story in full, and Chapter 4 follows Solid Edge and SolidWorks down their two very different roads to the same kernel.

Why a PLM person should care

You do not need to write kernel code. You do need to understand three consequences of the kernel's existence, because they show up in every integration project and every vendor selection.

Interoperability is a kernel problem, not a file-format problem. STEP does carry B-rep geometry. But each kernel has its own tolerancing rules and its own way of describing faces, edges and blends, so the receiving kernel has to rebuild the model under different rules. The rebuild fails or approximates most often at fillets, surface blends and near-tangent edges, which is exactly where manufacturing cares most. Exchange between two products on the same kernel is close to lossless. Cross-kernel exchange is where geometry silently degrades, and no file format fixes that. The CAD interoperability explainer goes deeper, and Chapter 14 looks at what USD and AI translators are doing about it.

Lock-in is structural, not contractual. A vendor practically cannot swap the kernel under a shipping product. Every feature, every constraint behavior and every legacy model depends on the engine's specific quirks. Autodesk forked ACIS into ShapeManager rather than move Inventor off it. SolidWorks never left Parasolid despite its parent owning CGM. Once you understand that, vendor roadmaps and M&A start to make sense: Siemens buying UGS and Dassault buying Spatial were kernel acquisitions first and product acquisitions second.

The new entrants are routing around the B-rep kernel. Cloud CAD like Onshape simply licenses Parasolid and competes on architecture. Implicit and signed-distance-field modelers like nTop are not B-rep at all, which is why they handle lattices and additive geometry that a classical kernel chokes on. Generative and AI-driven tools treat the kernel as a service to be called, not a product to be built. The CAD modeling paradigms article maps those three worlds, and the Best CAD Software 2026 guide shows which vendors sit in each.

How we got here

The kernel is the convergence point of fifty years of geometry research. Constructive solid geometry gave us Booleans on primitives. Faceted models gave us something to draw. Exact surfaces and then NURBS, descended from Bézier's work at Renault, gave us curves a machinist could trust. Parametric modeling put a feature history on top. Meshing and generative design sit downstream, consuming what the kernel produces. Chapter 2 traces the Cambridge origins that both Parasolid and ACIS share. Chapter 6 follows the parametric-to-hybrid shift, and Chapter 8 covers the surfacing technologies that made Class-A design possible.

Go deeper: the Kernel Wars series

The full table of contents has all fifteen chapters, with audio versions on the Listen page. If you only read three:

• Chapter 1: Graphics Kernel Anatomy 101, the layered model above in detail
• Chapter 3: Proprietary versus Licensed Kernels, the business decisions that shaped the market
• Chapter 15: The Kernel Wars, a Modern Perspective, where it is all heading

Then test yourself on the Kernel Wars quiz.

Related Concepts

• Geometry Kernel and CAD Kernel in the glossary, the formal definitions used across this site
• B-rep (Boundary Representation), the representation every mainstream kernel is built on
• NURBS, the surface math kernels rely on for freeform geometry
• What is CAD Interoperability?, why geometry degrades when it crosses kernel boundaries
• CAD Modeling Paradigms: NURBS, Parametric and Implicit, the three geometry worlds and which tools live in each
• Best CAD Software 2026, which platform, and therefore which kernel, fits which program

Frequently Asked Questions

What is the difference between a geometric kernel and a CAD system?

The kernel is the math library that creates and edits exact 3D geometry. The CAD system is everything wrapped around it: the sketcher, the parametric feature tree, the assembly and constraint logic, the viewport and the CAM or simulation modules. Several competing CAD products run on the same kernel, so the differences you notice between them live in those outer layers, not in the geometry engine.

Which geometric kernel does each major CAD system use?

Siemens NX, Solid Edge, SolidWorks and Onshape run on Parasolid, owned by Siemens. CATIA V5 and 3DEXPERIENCE run on CGM, Dassault Systèmes' own kernel. PTC Creo runs on Granite, PTC's own engine. Autodesk Inventor and Fusion 360 run on ShapeManager, Autodesk's fork of ACIS. AutoCAD and BricsCAD run on ACIS, sold by Spatial, a Dassault Systèmes company. FreeCAD runs on the open-source Open CASCADE kernel.

Why does the kernel matter for CAD interoperability?

Every kernel represents boundary geometry and topology with its own tolerancing rules and its own way of describing faces, edges and blends. STEP and IGES carry that geometry between systems, but the receiving kernel has to rebuild it under different rules, and the rebuild fails or approximates most often on fillets, surface blends and near-tangent edges. Exchange between two products on the same kernel, for example a Parasolid X_T file from SolidWorks into NX, is close to lossless. Cross-kernel exchange is where geometry silently degrades.

Can a CAD vendor switch to a different kernel?

In practice, no. Every feature, every constraint behavior and every legacy model in a mature CAD product depends on the kernel's specific behavior, so a swap is a multi-year rewrite with a migration problem for every customer file. That is why SolidWorks, a Dassault Systèmes product, still runs on Siemens' Parasolid, and why Autodesk forked ACIS into ShapeManager in 2001 rather than move Inventor to another engine.

Is a geometric kernel the same thing as a graphics engine?

No. The kernel holds the exact mathematical geometry. The graphics or visualization layer tessellates that geometry into triangles and hands them to OpenGL, Vulkan or a toolkit like HOOPS to draw on screen. What you see in the viewport is an approximation; what the kernel stores is the truth that CAM, simulation and PLM systems depend on.

How is AI changing geometric kernels?

Three ways. Generative design tools are heavy kernel clients, querying the engine thousands of times per study, so kernel API performance now matters commercially. A new class of implicit and signed-distance-field modelers sidesteps B-rep entirely for lattice and additive geometry. And machine-learning translators are starting to recover design intent across kernel boundaries, which is the subject of Chapter 14 of the Kernel Wars series.