Floating-point addition discrepancy between CUDA and OpenCL

Question

I\'m computing the following expression:

726.9997558594 + 3600

in a CUDA kernel, and in an OpenCL kernel, both running on the the same (NVIDI