The authors implemented an explicit GPU-based solver within the material point method framework and tested using two- and three-dimensional problems. Results seem to agree with the expected values validating this MPM - GPU architecture. I would like to suggest the publication of this work. Nonetheless, some minor points should be addressed first.

1. The authors mentioned that this GPU architecture speeds up information transfer between nodes and material points. As stated, this is one of the most computationally expensive operations in MPM. Nevertheless, finding material points new location after the mesh returns to its original position is another process that is computationally expensive (in many cases more expensive than information transfer between nodes and material points). I would like to know if the GPU architecture proposed also improves this step. If yes, the author could indicate it in the paper. If not, it would be interesting if the author discusses the possibility of combining some techniques (e.g. Pruijn N.S. 2016) together with GPU's to improve MPM computations.

Pruijn N.S. 2016. The improvement of the material point method by increasing efficiency and accuracy. TU Delft Master Thesis.

2. The authors include a damping value D in the simulations. This value is not well validated. It seems that several simulations were needed to find it, giving the idea that D is not related to the material properties and geometry and is more of an artificial way to reach the desirable results. The authors should elaborate better on the reasons for using this specific damping value.

Besides the previous comments, other minor issues should be reviewed.

1. In line 32, the authors mentioned that "The background mesh can be reset". As far as I know, the background mesh must be reset. I recommend changing the verb "can" for a better one.
2. I am wondering if the variables in line 75 are the same as in equations 2 and 3 since different punctuations were used (e.g. Å« and û). 
3. Finally, I recommend reading the paper again to correct some typos detected.

This paper demonstrates the numerical implementation of the explicit material point method using GPU for parallel computing. In Reviewer opinion, the paper is of interest for the Readers of Geoscientific Model Development. However, a number of issues need to be addressed before the paper can be accepted for publication.

1. The paper is currently long and several part can be replaced by references. For example, the MPM algorithm is derived from forward-Euler scheme with update stress lass, GIMP basis functions in appendix A.

2. There are several methods dealing with the volumetric locking in the literature. However, the author proposed the volumetric locking by averaging only the volumetric part of the stress tensor. The author is suggested to clarify the decision for that. Furthermore, volumetric locking can smooth out the value of the stress. It is better if the stress is plotted in the numerical examples to see the difference between simulations with and without volumetric locking.

3. Section 4 mentions that Model 1b demonstrates the influence of mesh resolution but I do not see it in Model1b. The author is suggested to perform convergence rate analysis in different mesh size in the plane strain to highlight the influence of the mesh resolution.

4. For Model 1b, the presented final geometry of the experiment is shorter to the one in Bui et al. (2008) experiment (see Figure 6) in my opinion. Please check. Therefore, it is not necessary to introduce the damping.

5. In Model 2, there is a boundary effect on the failure mechanism as the shear band can touch the bottom boundary. It would be better if there is a larger depth in the bottom direction. And, the Model 2 introduces local damping which in my opinion it is not necessary.