STAR-CCM+ CPU vs GPU Runtime Benchmarks with MayaHTT

Harnessing GPU Acceleration for STAR-CCM+

In the Computational Fluid Dynamics (CFD) field, speed and accuracy are the most important metrics. Engineers often struggle to find a happy balance between the two. Reducing the time to completion often means sacrificing model fidelity and resolution thus the overall accuracy of the model.

Siemens STAR-CCM+ has implemented and improved upon its GPU-accelerated computing support in the past couple of years. Engineers are adopting GPU acceleration to speed up their simulation times while maintaining model complexity.

Exxact and MayaHTT, a partner and ISV for STAR-CCM+ have collaborated to run some GPU runtime benchmarks to showcase the performance GPUs can offer over traditional CPU-only deployments. MayaHTT’s engineers ran their models on Exxact hardware to deliver these results.

Why GPU Acceleration Matters

The evolution of GPU computing has reached a point where performance matches CPU capabilities and, in many cases, surpasses them. For engineers and researchers, this means the ability to iterate faster and model larger simulations.

• Time: Less time to completion means more simulation tests run to fine-tune before production. Less time also means less time spent waiting for results, accelerating the workflow process.
• Complexity: Increased speedup even on highly complex models enables better representations of the simulation. While GPU VRAM is limited compared to CPU RAM, NVIDIA continues to release GPUs with increased VRAM capacities specifically for visualization and HPC workloads. The NVIDIA RTX 6000 Ada tested has 48GB of VRAM.
• Cost: A system featuring four NVIDIA RTX 6000 Ada has the equivalent performance of hundreds of CPU cores. However, hundreds of CPU cores require multi-node setups increasing total cost to run and maintain whereas a four GPU system can be workstation can sit on your engineer's desk or a single 2U server in a rack.

NVIDIA RTX 6000 Ada STAR-CCM+ Benchmarks

Let’s get into the benchmarks that the team at MayaHTT ran. We first had a benchmark on the classic Le Mans car which is an external aerodynamics simulation with coupled flow, steady flow, and RANS turbulence. We tested one large and one small case.

GPUs are likely not fully saturated in the 1.27 million cell simulation. For smaller cell count simulations, the GPU cannot effectively ramp up and showcase performance prowess with the time to completion in less than 5 minutes. However, on the larger 16 million cell count model, 4x NVIDIA RTX 6000 Ada delivers over 7.4x speedup.

We are also quite confident that 4x GPUs on an external aerodynamics simulation is likely also not fully saturating all the GPUs. Therefore, we decided to run a more intensive simulation with increased physics difficulty. The Flamelet Combustion simulation (SANDIA Flame D) includes flamelet combustion, segregated flow, reaction modeling, implicit unsteady, and LES turbulence.

Here we can see a more likely representation of GPU acceleration versus CPU runtime. GPU-accelerated computing delivers over 6.2x speedup versus CPU-computing. An 18M-cell Flamelet Combustion simulation solve times were cut from a week to a single day with 4x NVIDIA RTX 6000 Ada GPUs!

Key Takeaways

MayaHTT kept the configuration static during their benchmarks. When considering smaller scale simulation models (1.27M cell, 4M cell, and other similar size use cases), 4x GPUs are not the optimal configuration for the price to performance; the GPUs are not fully saturated and likely a single GPU will perform the same. However, for larger and more complex simulation we can expect a high degree of simulation speedup. Compared to 32 modern CPU cores from the Intel Xeon Scalable 6444Y, we can expect over 6x the performance in these use cases with 4x NVIDIA RTX 6000 Ada GPUs.

Another key thing to consider for your deployment is if your simulation requires a high degree of precision. The NVIDIA RTX 6000 Ada GPU is optimized for single-precision FP32 calculations and does not have native double-precision FP64 capabilities. Some other GPUs to consider that have native double-precision FP64 capabilities include:

• NVIDIA H200 NVL 141GB
• NVIDIA H100 NVL 94GB
• NVIDIA A800 40GB Active

The NVIDIA H200 and H100 are data center GPUs that can only be outfitted in a server and deployed in a rack. However, these GPUs can deliver performance equivalent to multiple hundreds of CPU cores. The NVIDIA A800 40GB Active is based on the NVIDIA A100 data center GPUs but serves as a workstation-class GPU with an active cooler.

Industries ranging from aerospace to automotive to civil engineering can benefit from incorporating GPU-accelerated computing into their computing infrastructure, reducing time-to-market while improving product quality. The ability to simulate larger, more detailed models also opens doors for innovation.

GPUs are not cheap but they are drastically worth time saving investment. To gain equivalent performance to a GPU with CPU-only hardware extends past the cost of the processor; it includes the maintenance, networking, system memory, and whole new systems for managing everything. Time is money and the ability to not only run subsequent simulations but run them faster dramatically improves business workflow and quality. A multi-node CPU cluster is costly to run and maintain versus a single GPU-accelerated system.

If you're considering upgrading your hardware for CFD workflows, leveraging GPUs is no longer an option—it's a necessity. If your workflow can take advantage of GPU acceleration, invest in your computing as more and more solvers continue to migrate and adopt GPUs. Contact us to learn how Exxact can tailor GPU-powered systems to meet your CFD needs. Thanks to MayaHTT for running these benchmarks on Exxact hardware.