A GPU based parallel computing mechanism of an etching profile evolution model
Since the simulations for etching and deposition are increasingly time consuming, we explore the parallel computing mechanism of a typical profile evolution model on GPU platform in order to accelerate the process. It turns out the traditional parallelism will lead to either data unsafety or low efficiency, because of GPU's special architecture and a lack of safe protocols to protect a memory address from being simultaneously written by multi-threads. To overcome the dilemma, we propose an advanced parallel computing mechanism suitable for GPU platform by dividing each Thread's work into several parts. It is theoretically and experimentally proven that the multi-Threads writing the same address can be totally avoided. Experiment results show that the advanced parallel model is much faster than the traditional serial one.
Full Text: PDF