CORe's basic output grid is the 512x256 Gaussian grid.  This is
the grid used by the data assimilation.

This is the layout in the grib files

grid_template=40:winds(N/S):
	Gaussian grid: (512 x 256) units 1e-06 input WE:NS output WE:SN
	number of latitudes between pole-equator=128 #points=131072
	lat 89.462947 to -89.462947
	lon 0.000000 to 359.296875 by 0.703125

The data is arranged WE:NS, So the first grid point is at 89.462947N and 0E.
The second grid point is at 89.462947N and 0.703125E.  For some software, 
grid layout is automatically converted to its standard layout.  For example,
wgrib2 converts the data to a WE:SN layout.

There are 256 Gaussian latitudes which do not include the poles. The latitudes
are not equally spaced.  The longitudes start at 0 and increment by 360/512 of a degree
eastward.

Why is the data assimilation done on a Gaussian grid?  Probably because the
model evolved from a spectral model where the model physics and data assimilation 
were done on the Gaussian grid. Converting the data assimilation from a Gaussian
grid to a cubed sphere grid would be an involved undertaking and wouldn't be
available in the early versions of the FV3GFS.