When using simd_inverse functions, the same input yields different results on different devices.
On the Mac mini with M4 pro and M3 ultra, the result is
But on the Mac mini of M2 ultra, the result is
What's wrong with simd_inverse?
Recommended
Hi,
We don't guarantee absolute numerical consistency across different devices. There are many hardware and software factors that can all introduce subtle differences.
That is accurate and appreciate the helpful reply. ^
As suggested there are a lot of factors with significant impact e.g. pro versus ultra architecture, number of cores, software differences etc.
In Metal, if you need greater precision you can disable fast-math but that doesn't necessarily guarantee lower variance between devices.