Accelerate multiphase CFD with GPU-native Volume Of Fluid (VOF) and Mixture Multiphase (MMP) solvers
- 4 hours ago
- 5 min read
Graphics Processing Units (GPUs) consist of thousands of identical cores, each designed to operate independently on massively parallel tasks, which can be subdivided so that each core works independently. This design differs from that of the traditional Central Processing Unit (CPU), which is composed of a smaller number of highly complex cores, with sophisticated control logic, large amounts of hierarchical cache memory, and advanced mechanisms such as out-of-order execution, branch prediction, and deep pipelines. This architecture is optimized to minimize latency in the execution of sequential tasks and to handle diverse and dependent instruction streams. GPUs, on the other hand, are designed to keep data and processing as local as possible within their multiprocessors, reducing data movement, increasing throughput, and maximizing performance in highly parallelizable workloads.
Computational Fluid Dynamics (CFD) is ideal for GPU architecture because everything is done locally in the computational cell, eliminating the need for communication with distant cells. Reducing the distance between electrons brings three main benefits: simulations can be run faster; power consumption per simulation is much lower; and the hardware footprint is much smaller.
The ability to run Simcenter STAR-CCM+ simulations on GPUs is not new, but Simcenter STAR-CCM+ 2602 took a major step forward by adding Volume of Fluid (VOF) and Multiphase Mixing (MMP) to the list of native GPU solvers. The multiphase capability supported on GPU in this version is impressive, including phase change models such as evaporation, boiling, and cavitation, acceleration techniques such as implicit multi-steps, and support for multiple regimes with MMP-LSI. Here are some examples of the benefits that GPU execution can bring to a variety of applications.
Run faster tank sloshing simulations
Solver: VOF;
Mesh: Uniform static mesh of 5.6 million pixels (AMR not yet compatible with GPU);
Time step: 5e-4s with dynamic substeps (target CFL 0.5);
Motion: Sinusoidal lateral motion;
CPUs used: 192 CPU cores (AMD EPYC 7532);
GPU used: 1 NVIDIA RTX 6000 Ada
The first example is a case of liquid oscillation in an automotive fuel tank. As engineers, we want to know how the center of gravity shifts as the tank oscillates, due to the loads it transfers to the vehicle, and the effect this will have on the vehicle's stability and dynamics. Liquid oscillation is also a concern in cryogenic applications, where boiling occurs frequently, which is also addressed in this version.
In Simcenter STAR-CCM+, the same solver was used for both CPU and GPU versions, meaning that if the cases converge well, identical results can be expected. In the tank oscillation example, this is exactly what is observed, with the free surface motion over time being almost identical (as in real experiments, VOF transient cases are stochastic in nature and therefore no run will be completely identical to the point where every drop coincides). The center of gravity motion shows good agreement with the experiment, both in CPU and GPU runs.
CPU
GPU
The advantage of running the program on the GPU becomes more evident when execution times are compared. A single GPU was significantly faster than 192 CPU cores. In fact, it would take 251 CPU cores to match the GPU's speed (a metric known as CPU core equivalence).
When comparing power consumption, the benefits of the GPU are clear, as it uses only 19% of the CPU equivalent, reducing operating costs and carbon footprint.
Speed up propeller cavitation simulations
Solver: VOF plus Schnerr-Sauer cavitation model;
Mesh: 4.4 million clipping static mesh (focused on the region near the propeller);
Time step: 5e-6s with 3 volume fraction substeps;
Motion: MRF;
CPUs used: 160 CPU cores (Intel Xeon Gold 6248);
GPU used: 1 NVIDIA Tesla V100
The next example is a marine propeller operating in a condition where cavitation is expected. This gives us the opportunity to test some of the advanced physics features included in the GPU VOF in this version. In this case, the Schnerr-Sauer cavitation model was used. The model predicts the growth and collapse of vapor bubbles due to low pressure on the propeller surface. These bubbles coalesce to form larger vapor pockets that fill the tip vortex and move downstream, forming a classic helical pattern. The results of this simulation on CPUs and GPUs are shown below. They are identical, as expected.
CPU
GPU
The single GPU completed the execution in about 70% of the time it would take for the 160 CPU cores, which is equivalent to 231 CPU cores. As in the previous example, the energy consumed to complete the execution is also much lower, with the GPU consuming only 35% of the energy used by the CPUs.
Accelerate marine resistance predictions: Kriso Container Ship (KCS)
Solver: VOF plus VOF waves
Mesh: 28M clipped static mesh
Time step: 0.02s
Movement: None
CPUs used: 512 CPU cores (AMD EPYC 7532)
GPUs used: 2 NVIDIA RTX 6000 Ada
Still on the topic of maritime applications, the next simulation is a drag calculation for the Kriso container ship (KCS) test case. Accurate drag prediction in these examples requires precise capture of free surface waves both around the vessel and downstream. This simulation is possible on GPUs thanks to VOF wave support in this version.
CPU
GPU
Once again, the CPU and GPU results are indistinguishable. Comparing execution time, both GPUs were slightly slower than 512 CPU cores, resulting in an equivalent of 214 CPU cores. The GPU's power consumption was only 30% of the CPU cluster's consumption.
Run E-Motor cooling studies faster
Solver: MMP-LSI;
Mesh: Static polyhedral mesh of 4.16 million iterations;
Time step: Adaptive time step with a maximum CFL target of 2 and 10 substeps;
Motion: Rigid body motion (with intersection based on metrics and distance from the PDE wall);
CPUs used: 160 CPU cores (Intel Xeon Gold 6248);
GPU used: 1 NVIDIA Tesla V100
The last example is an electric motor similar to those found in electric vehicles. These motors require cooling with a dielectric fluid (oil) which, in this motor, is injected through fixed inlets on the top of the machine over the copper windings. Optimizing cooling in an electric motor is fundamental to maximizing performance and efficiency. This simulation uses Multiphase Mixture Modeling (MPM) with Large Scale Interface (LSI) to allow the coexistence of resolved jets and dispersed mixtures of sub-grid droplets. The simulation also includes relative motion (Rigid Body Motion with sliding interfaces).
CPU
GPU
The results again show excellent agreement between CPU and GPU execution. In this example, the single GPU was slightly slower than the 160 CPU cores, resulting in an equivalent of 124 CPU cores and power consumption equivalent to 65% of that of the CPUs. This is not as good as in the other examples due to the need to re-intersect the sliding mesh at each time step (this is a non-local operation and therefore less suitable for GPUs). Even so, it still represents a very significant speed gain.
Take your multiphase simulations to a new level of speed and efficiency with the power of GPUs in Simcenter STAR-CCM+. CAEXPERTS can help you implement, optimize, and extract maximum performance from this technology in your projects. Schedule a meeting with our experts and discover how to accelerate your results, reduce computational costs, and innovate with greater confidence.
WhatsApp: +55 (48) 98814-4798
E-mail: contato@caexperts.com.br
