Chip simulations on NVIDIA GPUs sped up by as much as 35 times

NVIDIA has gathered in a single blog post several industrial deployments of its libraries in engineering software for semiconductor manufacturing. The most practical of them concerns more than just microelectronics. In COMSOL Multiphysics 6.4, the NVIDIA cuDSS GPU solver is built into the standard set of solvers and works for any single- or multiphysics model. Existing simulations can be moved to the graphics card without changing the physics settings, and for semiconductor problems COMSOL gains a speedup of up to 10 times, according to NVIDIA.

The other partners quote higher figures. Siemens is integrating NVIDIA cuLitho into the Calibre lithography suite and getting up to a 17-fold speedup on key computations. Applied Materials reports a 35-fold speedup of topography simulations in its ACE+ platform. Advantest, using NVIDIA AI agents, has cut response time at a test site from hours to minutes.

COMSOL moves the solver to the graphics card for all licenses

In the finite element method, most of the time goes into repeatedly solving large sparse systems of equations. Such systems arise in implicit time integration, in nonlinear solver iterations, in eigenfrequency analysis and in parametric sweeps. As Tyler Donegan writes in the COMSOL blog dated October 14, 2026, the cuDSS direct sparse solver accelerates exactly this stage thanks to GPU memory bandwidth and parallelism. The gain is therefore likely to be the more noticeable the more of these repeated solves a problem contains.

Take an engineer tuning the parameters of a process chamber and running a sweep across dozens of geometry variants with heat, electrics and mechanical stresses. Previously they had to choose between simplifying the model and overnight CPU runs. Now the same project can be launched on the graphics card without rebuilding the model. COMSOL lists the conditions outright:

• the GPU solver and the training of surrogate neural network models on the GPU are part of the standard installation for all license types;
• all you need is a compatible NVIDIA graphics card and drivers;
• time-explicit pressure acoustics computation on the GPU is available with the Acoustics Module.

The "up to 10 times" figure is a generalized estimate from NVIDIA's material. In COMSOL's own tests on semiconductor models with an NVIDIA RTX Pro 6000 Blackwell Max-Q 96 GB graphics card, the gain compared with proven CPU configurations ranged from a little over two times to 13 times depending on the problem. The real benefit depends on what share of the computation is taken up by solving the linear systems.

Applied Materials accelerates the chain from atoms to the shop floor

Applied Materials is plugging NVIDIA libraries in at several levels at once. Its Ginestra platform with cuDSS runs up to 10 times faster than on processors alone. NVIDIA's technical blog dated July 26, 2026 specifies that this makes it possible to run thousands of virtual experiments on material combinations and reaction pathways. The same post cites a measurement for density functional theory calculations with the cuEST library: a problem that took five days on 64 CPU cores is solved in about two hours on a single NVIDIA B200, roughly 55 times faster.

In the ACE+ platform the speedup is achieved differently. The PhysicsNeMo framework turns multiphysics simulation results into digital twins that run in real time. Topography is modeled up to 35 times faster, and parameters can be explored interactively. Lam Research has shown a similar approach: according to NVIDIA, PhysicsNeMo predicts plasma behavior in etching processes in milliseconds, whereas a full simulation takes days.

The top level of this chain is digital twins of entire fabs built on the Omniverse libraries. In them, Applied Materials models the movement of materials, looks for bottlenecks and compares layouts and operating strategies before anything changes on the real line.

Lithography and testing

In computational lithography, that is, in calculating how a circuit pattern is transferred onto silicon, NVIDIA speaks of problems too heavy for CPUs at production scale. Integrating cuLitho into Siemens Calibre delivers up to a 17-fold speedup of key workloads. According to NVIDIA, this widens the range of what engineers can check before printing a wafer. Exactly which tasks and what hardware produced that figure is not disclosed in the materials.

Advantest uses NVIDIA NeMo Switchyard agents to pick an AI model for a specific task. At a test site this used to be a sequential manual process that took hours, and now the answer arrives in minutes. The materials contain no details about the process itself or about the NeMo Switchyard license.

NVIDIA frames the overall idea this way:

"The future of semiconductor manufacturing lies in uniting these capabilities: production data feeds the models, the models evaluate changes, and the conclusions return to the engineers who make manufacturing decisions."

Who will find this useful beyond the fabs

For most engineering readers, the main news in this package is COMSOL. The solver works with any multiphysics model, including problems outside microelectronics, and does not require a separate license. For the owner of a workstation with a suitable NVIDIA graphics card, judging by the description, it is enough to update to 6.4 and compare times on their own problem. That said, the acceleration is built on CUDA and designed for NVIDIA graphics cards.

The figures from Applied Materials, Siemens and Lam Research relate to closed industrial packages, and they are hard to verify from the outside. The approach with physics-trained models, as in ACE+ and in the Lam Research plasma example, has its own price. The fast answer is probably only possible after training on the results of full simulations, so full computations are still needed for that. The 35-fold speedup apparently applies to exploring parameters within an area that has already been studied; computing new physics from scratch still requires a full simulation.

Whose GPU acceleration result impresses you more?

• NVIDIA: up to 10 times
• Siemens: up to 17 times