Quoted from: https://geodynamics.org/cig/software/specfem3d_globe/ 

Effects due to lateral variations in compressional-wave speed, shear-wave speed, density, a 3D crustal model, ellipticity, topography and bathymetry, the oceans, rotation, and self-gravitation are all included.

The version 7.0 release offers GPU graphics card support for both OpenCL and CUDA hardware accelerators, based on an automatic source-to-source transformation library (Videau et al. 2013). It offers additional support for ADIOS file I/O formats and contains important bug fixes related to 3D topography and geographic/geocentric transformations. Seismogram file names adapt a new naming convention, with better compatibility to the seismogram specifications by the Incorporated Research Institutions for Seismology (IRIS).

The version embeds non-blocking MPI communications and includes several performance improvements in mesher and solver. It provides a perfectly load-balanced mesh for 3D mantle models honoring shallow oceanic Moho (depths less than 15 km) and deep continental Moho (depths greater than 35 km). It also accommodates European crustal models EPcrust (Molinari & Morelli, 2011) and EuCrust07 (Tesauro et al., 2008), which may be combined with global crustal model Crust2.0. Sedimentary wavespeeds are superimposed on the mesh if sediment thickness exceeds 2 km.

Additional new model routines are provided for the Comprehensive Earth Model (CEM) project, generic point-profile models (PPM) and Gauss-Lobatto-Legendre based models (GLL), with complementary tools for postprocessing adjoint sensitivity kernels and gradient-based model updates. The structure of the software has been simplified to facilitate easier implementation of new 3D models. The code accommodates general moment tensor files, and provides complete information in the SAC headers, as explained in detail in the updated user manual. New matrix-matrix multiplication routines, adapted from the book of Deville et al. (2002), and loop-vectorization help reduce the total number of memory accesses performed in each spectral element and improve code vectorization, thus enhance numerical performance of the version.