cublas sgemv accuracy

The accuracy of the Cublas_sgemv function seems to decrease as the matrix dimensions increase.

For a tiny 2x2 matrix, the result is perfectly accurate.
For a larger 10x100 matrix, the result is accurate to about 5 sig figs.
For a much bigger 100x32560 matrix, the result is in the right order of magnitude, but every sig fig is wrong!

Has anyone else experienced this problem? Has a solution been found or on the way?

Thanks very much for any input.