You should not expect the exact same results between CPU and GPU for floating point arithmetic.
CPU and GPU have different instruction sets. When you compile for GPU, the order of operations can be different than for CPU, including because of differences in the optimization capabilities of the different compilers. For efficiency, a number of operations on GPU use vendor high-performance libraries, which use different implementations than the high-performance libraries used for CPU. Whether the high performance libraries are even invoked on CPU depends on the problem size (there is a cost to starting up those library calls, so MATLAB does the calculations more directly for smaller calculations.)
