CFD Simulation for Large-Scale Modeling
In the rapidly evolving field of computational fluid dynamics (CFD), the ongoing debate between CPU and GPU performance is becoming more pronounced. CFD simulation has evolved significantly, moving from simple 2D simulations to complex, large-scale 3D models that engineers rely on today. With advancements in both hardware and software, professionals can tackle increasingly complex problems with enhanced accuracy. However, despite these technological improvements in high-performance computing (HPC), challenges persist, particularly for large-scale simulations that require considerable computational power and time.
Traditional CPU-based solvers, while powerful, often struggle with the expansive computational needs of modern CFD simulations. As simulation models grow in detail, mesh sizes increase, and more physical phenomena are incorporated, the limitations of CPU solvers become more apparent. This situation underscores the necessity to explore alternative computing solutions, particularly those involving GPU computing, to meet the demands of sophisticated CFD analysis.
GPU Ansys Fluent Solver: A Game Changer!
CFD simulations have traditionally relied on CPU processors, and the Ansys Fluent solver effectively scales with an increasing number of cores. However, this method can be both time-consuming and energy-intensive, especially for large and complex simulations. Enter GPU computing, an innovative approach that has the potential to revolutionize the field of CFD. With higher-bandwidth memory and lower energy requirements for executing instructions, GPUs enable higher fidelity modeling and significantly faster simulations. This advancement allows engineers to test multiple designs and conditions simultaneously, ultimately streamlining the CFD process and enhancing efficiency.
As shown, the Ansys benchmark illustrates a 91x performance boost in the CFD analysis of a sedan car with 4 million mesh elements, thanks to CPU advancements from 8 cores in 2008 to 112 cores in 2023. Over these years, we've seen a 14X increase in cores, translating to about a 6.5x improvement in CFD analysis performance per core. However, the Fluent solver can leverage native GPU computing, achieving CFD simulations that are about 9X faster with a single H100 GPU card compared to 112 CPU cores. This significant boost in performance and efficiency highlights the potential of GPUs to address the computational challenges faced by large-scale CFD simulations.
Ozen Engineering: Hardware and CFD Test Model
In a series of tests conducted using Ansys Fluent, we ran steady-state conjugate heat transfer simulations with approximately 16 million mesh elements. The CFD model and boundary conditions are illustrated below.
For these tests, we utilized the AMD Epyc 7543 and Intel Xeon Gold 6242 processors from the HPC-Corvid Cloud System. The AMD Epyc processor is equipped with 64 cores/node and 250 GB RAM/node. The Intel Xeon Gold processor has 32 cores/node and 190GB of RAM/node. Additionally, we employed an NVIDIA RTX 6000 card 51.5GB RAM from Ozen Engineering for GPU computation. This setup allowed us to evaluate the performance of different hardware configurations in handling large-scale CFD simulations.
For the test cases, we utilized the workstations employing multiple cores ranging from 16 to 128. We recorded the wall-clock time required to achieve a converged solution, as detailed in the tables below. The AMD processor demonstrated faster performance when all cores were utilized from a single computation node, possibly due to larger RAM on the node. However, the Intel processor slightly outperformed the AMD processor when the Fluent simulation was run across multiple nodes (for 96 and 128 cores).
Next, we executed the same conjugate heat transfer model under identical operating conditions using the Fluent native GPU solver with an NVIDIA RTX6000 card, as illustrated below.
Using the Fluent native GPU solver with an RTX 6000 card for the conjugate heat transfer problem is faster than using AMD and Intel CPU processors with up to 64 cores. The GPU native solver is approximately 2.5 times faster than AMD processors and 3.5 times faster than Intel processors when compared to a 16-core CPU processor. However, when utilizing more cores, such as 96 and 128, the CPU processors outperformed the GPU native solver in this case. To surpass the performance of a CPU with 96 and 128 cores, a more powerful GPU card, like an A100 GPU card, is necessary.
The runtime for CFD simulations may vary slightly depending on how the model is initialized before the simulation begins in Fluent.
Considerations for Using Fluent Native Solver
- In this benchmark study, all CPU simulations were conducted using double precision (dp), while GPU solver runs utilized single precision (sp), resulting in more iterations needed for a converged solution with the GPU solver. However, the results for temperature distribution inside solids and fluids are almost identical. When using a dp solver on the GPU, the startup time for the solution was significantly longer. Ansys also uses sp for Fluent benchmarks to reduce the GPU's memory usage.
- For simulations involving heavier meshes, high-performance GPU graphics cards like the A100/200 are necessary for quicker runs.
- In this study, the performance and runtime of the GPU native solver with the RTX6000 card remain almost unchanged when increasing CPU processors from 4 to 8 cores. However, more CPU cores are generally needed for a faster startup of the solution with the GPU solver, as they help distribute solution data more quickly from CPU cores to the GPU's numerous memory partitions.
-
The Ansys Fluent GPU Native solver is compatible with both AMD and Nvidia graphics cards for GPU computing.
.
Will GPUs Replace CPUs in CFD Simulations?
With the significant improvements in performance and energy efficiency offered by GPU computing, it's anticipated that GPUs might eventually take over from CPUs as the primary hardware for CFD simulations. Please check the references below. It's crucial to know that the GPU solver currently lacks support for the complete range of Fluent physics models. As GPU technology advances and each new Ansys release introduces more features, the likelihood of GPUs becoming the preferred hardware for engineering simulations increases. The future of CFD simulations could very well be led by GPU computing.
What types of industries can benefit the most from GPU Solver?
Industries that rely heavily on complex, large-scale simulations can benefit the most from GPU solvers. These include the aerospace, automotive, energy, and manufacturing sectors, where accurate and efficient simulations are crucial for design and development. While high-performance GPU cards come with a high cost, utilizing GPU computing allows these industries to achieve quicker solve times, lower energy consumption, and the capability to handle more complex simulations, ultimately resulting in superior products and innovations.
Ozen Engineering Expertise
Ozen Engineering Inc. leverages its extensive consulting expertise in CFD, FEA, optics, photonics, and electromagnetic simulations to achieve exceptional results across various engineering projects. Ozen Engineering leads in adopting advanced computational methods like GPU computing for high-quality CFD simulations. Equipped with expert teams and cutting-edge hardware, we help clients tackle modern CFD complexities, ensuring efficient and effective solutions.
We offer support, mentoring, and consulting services to enhance the performance and reliability of your systems. Trust our proven track record to accelerate projects, optimize performance, and deliver high-quality, cost-effective results for both new and existing water control systems. For more information, please visit https://ozeninc.com.
Suggested blog and video by Ozen Engineering
- The future of engineering computing - From workstations to the cloud: https://blog.ozeninc.com/resources/the-future-of-engineering-computing-from-workstations-to-the-cloud
- Ansys Fluent GPU solver speed test: https://www.youtube.com/watch?v=ynwtYF8QD1s
- The GPU Advantage: Lumerical's FDTD Simulation: https://blog.ozeninc.com/resources/the-gpu-advantage-lumericals-fdtd-simulation-breakthrough
- The future of engineering computing - From workstations to the cloud: https://blog.ozeninc.com/resources/the-future-of-engineering-computing-from-workstations-to-the-cloud
References
- Ansys Fluent GPU Solver FAQs: https://innovationspace.ansys.com/knowledge/forums/topic/fluent-gpu-solver-faq/
- Axial turbine simulation: https://investors.ansys.com/news-releases/news-release-details/ansys-baker-hughes-and-oak-ridge-national-laboratory-set-new
- Cardiovascular Research: https://investors.ansys.com/news-releases/news-release-details/ansys-and-nvidia-demonstrate-new-era-silico-cardiovascular
- Volvo cars simulations: https://investors.ansys.com/news-releases/news-release-details/volvo-cars-leverages-ansys-and-nvidia-gpus-accelerate-cfd
- Unleashing the Power of Multiple GPUs for CFD Simulations: https://www.ansys.com/blog/unleashing-the-power-of-multiple-gpus-for-cfd-simulations#:~:text=Why%20use%20GPUs%20for%20CFD,all%20costs%20are%20factored%20in.