- A PhD student develops a gaming technology that runs complex simulations on a GPU.
- The simulation involves crashing giant ocean waves against offshore wind turbines.
- A dambreak simulation on a GPU ran up to 4.5 times faster than 16-threaded CPU run times.
Alex Chow, who is pursuing PhD at The University of Manchester has built a program to perform complex engineering and scientific simulations on graphic processing units (GPUs).
So far, high-end graphic cards are used for creating realistic visuals and fast gameplay for PCs, laptops and gaming consoles. But now, GPUs are emerging as a technology to speed up complex simulations, executing several applications more than hundred times faster than traditional CPUs.
The ultimate goal is to perform large-scale simulations on graphic cards instead of supercomputer. Since supercomputer consists of hundreds of CPUs connected in parallel, they consume a lot of power while performing billions of calculations. Also, they are very expensive and are available to a small number of scientists and researchers.
On the other hand, GPUs are energy efficient and much cheaper than conventional supercomputers. They don’t need a whole room or exclusive facility. In fact, modern graphic cards are compact enough to be fitted in a laptop.
What Simulation Has Been Performed So Far?
Chow has developed a software that can create large-scale simulations of violent fluid flows on powerful graphic cards. The simulation involves crashing giant ocean waves against offshore wind turbines, to better examine the forces (including the potential of impact) exerted on the structures.
How Did He Do It?
The software is developed using an open-source code, named DualSPhysics, which is based on a smooth particle hydrodynamics (SPH) model. The code allows the complex simulation (like violent hydrodynamic flows) to run on a GPU. It can handle computation of millions of data points for 3D scientific applications on a single device.
For Chow, the most challenging part was solving mathematical systems of million of equations concurrently that rapidly change throughout a simulation.
The incompressible SPH is executed by optimizing the weaklycompressible SPH code and integrating it with ViennaCL (open source linear algebra library) for fast implementation of the Pressure Poisson equation (PPE).
A PPE matrix is created for moving particles at specific intervals, in order to optimize the limited memory of the GPU. The incompressible SPH pressure projection algorithm is executed at 4 different levels. Also, a precise and robust boundary condition is established for efficient parallel processing.
Flow diagram of key repeating steps in DualSPHysics Predictor–Corrector timestep on a GPU
Numerous validation cases are shown in this research to demonstrate the preciseness, speed and flexibility of the technology. For instance, a simulation of dambreak on a GPU ran up to 4.5 times and 18 times faster than 16-threaded and single-threaded CPU run times, respectively.
How This Simulation Can Help?
The United Kingdom produces 5 percent of annual electrical energy from offshore wind, which is expected to grow to 10 percent within next 2 years, and it’s growing worldwide.
Sometimes the ocean environment is extremely harsh and violent, which is why developing structures for them isn’t an easy job. Physical experiments on these environments would be very expensive and time consuming, or you can say it’s not practical.
These simulations will help scientists and engineers to take crucial steps and decisions about structure’s design and industrial free-surface hydrodynamic engineering applications, without investing in expensive experiments.