Numerical discrepancies when running scientific software on AMD64 machine on ARM

[AntoineGautier]

I use OrbStack to emulate AMD64 Ubuntu on ARM. I typically run scientific software and occasionally notice small numerical discrepancies compared to the same calculations performed on native Linux machines or even on WSL Linux running on AMD64 Windows. These programs rely on gcc for compiling auto-generated C code and linking mathematical libraries (libm, liblapack, libsundials). I can confirm that the version of gcc and these libraries are the same across the different environments.
I am curious about the origins of these numerical discrepancies. Could they be due to the emulation process or differences in the Linux kernel itself? Are there specific design choices in OrbStack that might explain them?

[kdrag0n]

Probably because of Rosetta not emulating 80-bit x87 FPU instructions accurately. We won't be able to do anything about this, but could you share a reproducer?

[AntoineGautier]

I'll try to find a reproducer using https://openmodelica.org/ and attach it here. (Previous occurrences of this issue relied on proprietary software.)