Why is GPU training often not bitwise reproducible?
PICTURE THIS: DATA SPLIT
Fit on train, tune on val, report on test once.
Simple meaning
Atomic adds, cuDNN autotune, and mixed precision can change reduction order and thus floats.
WHY — Reproducibility instead of guessing?
Why interviewers care about Reproducibility:
on Reproducibility.
the situation, the default choice, and one exception - that reads as experience.
Name the idea, why it exists, then one short example.
End with when you use it and one common pitfall.
STEPS — What happens step by step?
Before you speak the answer, walk the interviewer through these steps:
- 1Atomic adds, cuDNN autotune,
and mixed precision can change reduction order and thus floats.
- 2You can force deterministic
flags at a speed cost, but most teams accept close metrics plus logged environments.
- 3Interviewers want honesty here,
not a claim of perfect identity.
- 4Give an example
One tiny concrete case you can say aloud.
- 5Common mistake
What juniors usually get wrong.
- 6Close
When you pick this over the alternative.
EXAMPLE — See it in action
Here's a short line you can speak, broken into clear beats:
Note: Adapt this scaffold to your own project — keep it under 60–90 seconds.
Key takeaway
Atomic adds, cuDNN autotune, and mixed precision can change reduction order and thus floats. You can force deterministic flags at a speed cost, but most teams accept close metrics plus logged environments.