Carte graphique dans un ordinateur

CPU or GPU: which one runs the simulations?

Computers use two processors for calculations: the CPU and the GPU.

The CPU is the central processor. It runs the computer and almost all software. The GPU is the processor on the graphics card. Originally, it is used to display images and run video games.

The latest innovations in video games show instant physics simulation calculations: explosions, parts breaking apart, thanks to the power of GPUs. Yet in simulation, the same type of simulation takes hours, even days to calculate. Why not use the GPU to run the simulations, then?

Two ways of working

The CPU has a few dozen cores, but they are very powerful. Each core can do complicated calculations, one after the other, in the order you want.

The GPU is the opposite. It has thousands of smaller, simpler cores. Taken one by one, they are weaker. But they can all work at the same time, on the same kind of calculation.

It is a bit like comparing a few very fast specialists with a large team repeating the same task at the same time.

Why simulation still runs mostly on CPU

The software was made for the CPU. The simulation software used in engineering has been around for decades. It was written for the central processor and improved year after year.

The programming is not the same. Programming for a graphics card is not like programming for a central processor. The languages are not the same, and neither is the way of thinking about the calculation. For the GPU, the work has to be split into thousands of small calculations that do not depend on each other. Rewriting a whole software package takes years. That is why the migration happens in small steps: linear statics and simple contact move to the GPU first, while more advanced simulations stay on the CPU.

Many calculations are done in steps. In finite element analysis, the result of one step is used to calculate the next one. When the steps depend on each other, you cannot run everything at the same time. That is where the GPU loses its advantage.

The card’s memory is small. A large simulation model uses the computer’s main memory, which can have hundreds of gigabytes. A graphics card’s memory is much smaller. If the model does not fit entirely on the card, time is lost going back and forth, and the gain disappears.

Precision is expensive. Engineering software works in double precision, with about 16 significant digits per number. Video games work in single precision, with about 7 digits. Consumer graphics cards are very fast in single precision, but much slower in double precision. Professional cards that handle 16-digit precision exist, but the costs are significant.

The strengths and limits of each

The CPU does everything: all the physics, large models, with software proven over a long time. Its weakness: when the same calculation comes back millions of times, it is slower than a well-used graphics card.

The GPU is very fast when the calculation repeats and splits up well, like fluid flow, impacts or particles. In good cases, a calculation of several days finishes in a few hours. On the other hand, its memory is limited, the software has to be adapted, and many physics do not exist yet in a GPU version.

The trend: the two share the work

The trend is not the GPU replacing the CPU. It is the two sharing the work.

Publishers add the GPU piece by piece. At Ansys, the Fluent flow solver can run entirely on the graphics card, and Mechanical has offered a mixed CPU and GPU solver since 2025.

There is also software designed for the GPU from the start, like Ansys Discovery. The calculation is done directly on the graphics card and the result displays almost in real time when the model is modified. The results are instant, but it is more limited than the Ansys Mechanical solver.

Artificial intelligence pushes in the same direction: it runs on the same cards, so the cards keep getting more powerful, and simulation software benefits from it.

In the end, the CPU keeps the step-by-step calculations and the logic, and the GPU takes the big repetitive calculations. In our projects, we choose case by case, depending on the physics to simulate and the size of the model.

The difference in pictures

A picture is worth a thousand words. The following video is a Mythbusters demonstration that visually explains the difference between the CPU and the GPU.

Sources

Similar Posts