That's not a problem. The application in question was a redesign of an MPI FORTRAN implementation that can produce the same outputs to within FPRE. To verify the low-level code, we just run in double-precision and make sure it matches to 1 part in 10e-9 so across 200+ daily conformance tests. Production mode runs in single-precision with 64-bit Fixed Point accumulation. And that's also been demonstrated to have numerical stability nearly equivalent to running in full double-precision.