Adding on to Gang's comments, I think you have three basic ways to analyze this kind of data that can be mapped to the surface, and each of these has a multitude of variations in important details. I think your question has less to do specifically with 3dMVM than with the underlying methods.

1. Volumetric analysis. That is your @auto_tlrc method, but you can improve this with better volumetric registration to a standard space. auto_warp.py and @SSwarper can do this using nonlinear alignment. Your results will align better too across subjects than with the simpler affine only alignment of @auto_tlrc.

2. Volumetric analysis for the linear modeling, but statistical analysis on the surface. That is your second method. The correspondence across subjects is dramatically increased for this because the domain is limited with correspondence by nodes of the surface. This will follow the gyri and sulci more reliably.

3. Surface analysis for linear modeling. Here, you will likely find the best correspondence of these three situations because smoothing is done on the surface, respecting the topological boundaries better. Statistical analysis is still done on the surface. In this case, the EPI data is mapped to the surface after registration to an anatomical dataset. The volumetric data is mapped to the surface at that point, usually with an averaging through the cortex. Smoothing is then limited to the surface and the linear modeling is done then on the surface. This method is generally our recommendation for surface-based analysis. You can find examples in afni_proc.py's help for surface-based examples and in the AFNI class data under FT_analysis.

Note surface analysis is appropriate when you're interested, well, in the surface, namely the cortex, as the surface correspondence done by packages like FreeSurfer is based on the cortex. When you're interested in the rest of the brain, amygdala, striatum,..., then volumetric analysis is more appropriate. Of course, there's nothing stopping you from doing both. Regarding surface-based analysis, the data is originally volumetric, so there are choices on how to map data to the surface.