NG on GPU (I)

Computational Electromagnetics Lab, UC San Diego

Shaojing Li

Shaojing Li, Boris Livshitz, Vitaliy Lomakin

Here is a brief introduction to GPU, if you never heard of this term. At the same time, if you would like to know the reasons of doing computational science on GPU platform, please check this page and links in Related Links page.

In this project, we implement the already fast NG algorithm on nVidia GPUs. The implementation is not that straightforward for a complicated, multi-level algorithm such as NG and it is at initial stage that dealing only with static and low frequency field potential calculations. However, the results are very promising and exciting.

Non-uniform Grid Methods on Graphic Processing Unit (GPU)

*All images on this website are collected from the Internet. They are used for non-commercial purpose and all copyright information are kept. If you believe any of those images are used inappropriately, please contact the administrator of this website using the contact information provided in Contact section.

Last Update: Jun. 16th, 2012

Home

Research

Publications

Contact

Personal

Related Links

When it comes to low-frequency dynamic field calculation, the situation is more favorable to GPU. The reason is the computational burden is heavier in this case as a more complex Green's function is needed to describe the interaction between sources. Here is a figure on computational times of different number of unknowns. We can clearly see from this log-log figure that the computational complexity, which manifested here as computational time, scales as O(N²) for direct method and O(N) for NG algorithm. This leads to the astonishing speed up ratio as shown below.

The GPU platform we use is nVIDIA CUDA. It still has more potential and we will try to utilize it as much as possible in the near future. The question is no longer whether GPU is suitable for scientific computing and simulation of physical phenomenon, but how much better can GPU-based algorithm be. At least, it can bring the simulations that traditional requires clusters of hundreds of nodes down to a single workstation.

Back...

Table 1 The computational time comparison between different algorithms and implementations. Electro/magneto-static interactions between N sources only.

We can see from the table that the acceleration of provide by GPU is enormous.