Streaming is not suited for all applications or even all portions of an application.
Not all of the algorithmic tweaks that make RTM run efficiently on a CPU are feasible.
Some algorithmic complexities, such as the PML boundary condition and
variable grids, eat up too many FGPA
resources to be viable. The number of arithmetic units is finite. Approaches such as
multiple time steps per pass can oversaturate these resources. In addition, while
wavefield precision and compression can significantly improve
performance, their viability when using the RTM kernel for waveform inversion
needs to be investigated.
2009-10-16