I dived into pop grains and discovered that one of the nodes doing the heavy work is the get neighbors (Geometry Vop) node. Now as grains usually deal with constant number of particles, wouldn't it lend itself naturally to have the neighbors computed by Open CL?
Wouldn't that be faster than vex?

OpenCL doesn't know how to compute neighbors. That is why the vex function is used ahead of time with the neighborhood stored as an array.