As the examples demonstrate, although rounding follows deterministic rules, non-associativity introduces nondeterminism, which is further amplified by the larger rounding error of BF16. This becomes particularly relevant in the parallel computations performed in GPUs during LLM inference.

Ya might want to just read the paper...