next up previous [pdf]

Next: Acknowledgments Up: Yang et al.: GPU Previous: Conclusion

Discussion

It is important to point out that FWI can be accelerated in many ways. A good choice of preconditioning operator may lead to fast convergence rate and geologically consistent results (Ayeni et al., 2009; Guitton et al., 2012; Virieux and Operto, 2009). Multishooting and source encoding method is also a possible solution for accelerating FWI (Schiemenz and Igel, 2013; Moghaddam et al., 2013). These techniques can be combined with GPU implementation (Wang et al., 2011). There are many reports advocating their acceleration performance based on particular GPU hardware. These reports may be out of date soon once the more powerful and advanced GPU product are released. Although the speedup performance of our implementation may be a little poor due to our hardware condition, we believe that it is useful to give readers the implementation code to do performance analysis using their own GPU cards. The current GPU-based FWI implementation parallelizes the forward modeling process which makes it possible to run FWI on a single node and low-level GPU condition even for a laptop. However, it is completely possible to obtain higher speedup performance using the latest, high performance GPU products, and further parallelize the code on multi-GPU architectures using message passing interface (MPI) programming.


next up previous [pdf]

Next: Acknowledgments Up: Yang et al.: GPU Previous: Conclusion

2021-08-31