Can I expect exactly the same computation results with or without MIG use?

AI & Data Science NVIDIA Nemotron
1

Thank you for reading.

I am developing a foundation model (GPT-like model) using A100 80GB.
I get slightly different inference results when using MIG instances and when using the entire A100.
Are slight numerical errors expected when using MIG ?

(Note: I use multiple transformer blocks, and bfloat16 for inference)